[
  {
    "slug": "agent-skills",
    "title": "Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic",
    "url": "https://ai.engineer/talks/agent-skills",
    "videoId": "CEvIs9y1uog",
    "event": null,
    "durationMs": 982000,
    "speakers": [
      {
        "name": "Barry Zhang"
      },
      {
        "name": "Mahesh Murag"
      }
    ],
    "topics": [
      {
        "slug": "coding-agents",
        "name": "Coding agents"
      }
    ],
    "summary": "Barry Zhang and Mahesh Murag of Anthropic argue that instead of building domain-specific agents, developers should build reusable Skills—organized folders of files that package procedural knowledge for agents. They explain that skills are progressively disclosed to protect the context window, use scripts as self-documenting tools, and have already grown to thousands in five weeks, including foundational, partner, and enterprise skills. Skills complement MCP servers by providing expertise while MCP handles connectivity. The future includes treating skills like software with testing and versioning, and enabling agents to create their own skills for continuous learning, ultimately creating a collective knowledge base that makes agents more capable and reliable."
  },
  {
    "slug": "ai-coding-workflow",
    "title": "Full Walkthrough: Workflow for AI Coding — Matt Pocock",
    "url": "https://ai.engineer/talks/ai-coding-workflow",
    "videoId": "-QFHIoCo-Ko",
    "event": null,
    "durationMs": 5790000,
    "speakers": [
      {
        "name": "Matt Pocock"
      }
    ],
    "topics": [
      {
        "slug": "coding-agents",
        "name": "Coding agents"
      }
    ],
    "summary": "Matt Pocock presents a hands-on workshop on building a full AI-assisted coding workflow, arguing that software engineering fundamentals—not hype—make agents effective. He introduces the 'smart zone' and 'dumb zone' of LLMs (performance drops after ~100k tokens) and the 'Memento problem' (agents forget between sessions). His process starts with a 'Grill Me' skill that relentlessly questions the user until shared understanding is reached, then produces a PRD without reading it, slices work into vertical 'tracer bullet' issues, and runs agents AFK using TDD. He advocates designing codebases with deep, testable modules and shows Sandcastle, a TypeScript library for parallel agent execution with separate implementer (Sonnet) and reviewer (Opus). The workshop transforms ambiguous briefs into shippable features while keeping humans in the loop for QA and taste."
  },
  {
    "slug": "software-fundamentals-for-ai-coding",
    "title": "\"Software Fundamentals Matter More Than Ever\" — Matt Pocock",
    "url": "https://ai.engineer/talks/software-fundamentals-for-ai-coding",
    "videoId": "v4F1gFy-hqg",
    "event": null,
    "durationMs": 1106000,
    "speakers": [
      {
        "name": "Matt Pocock"
      }
    ],
    "topics": [
      {
        "slug": "coding-agents",
        "name": "Coding agents"
      }
    ],
    "summary": "Matt Pocock argues that software fundamentals matter more than ever in the AI era, directly countering the 'specs-to-code' movement that treats code as cheap. He shares practical skills like 'Grill Me' to reach a shared design concept with AI, a ubiquitous language from domain-driven design to align terminology, and test-driven development to force small, deliberate steps. Drawing on John Ousterhout's 'deep modules' and Frederick Brooks's 'design concept,' he advocates structuring codebases with simple interfaces behind which AI can implement freely. He warns that bad code is the most expensive it's ever been, as it blocks the productivity gains AI can offer. His reusable skills are available in the GitHub repo MattPocock/skills."
  },
  {
    "slug": "specifications-are-the-new-code",
    "title": "The New Code — Sean Grove, OpenAI",
    "url": "https://ai.engineer/talks/specifications-are-the-new-code",
    "videoId": "8rABwKRsec4",
    "event": null,
    "durationMs": 1296000,
    "speakers": [
      {
        "name": "Sean Grove"
      }
    ],
    "topics": [
      {
        "slug": "coding-agents",
        "name": "Coding agents"
      }
    ],
    "summary": "Sean Grove of OpenAI argues that specifications, not code, are becoming the fundamental unit of programming, with the most valuable skill being precise communication of intent. He presents OpenAI's Model Spec—a collection of versioned Markdown files—as a living specification that aligns both humans and models around shared values and intentions. Grove illustrates how the Model Spec served as a trust anchor during the GPT-4o sycophancy bug, where shipped behavior contradicted the spec's explicit 'don't be sycophantic' clause, leading to a rollback. He explains deliberative alignment, where the spec is used as training and eval material to embed policy into model weights, moving from inference-time prompting to muscle memory. Drawing parallels to the US Constitution, Grove positions specifications as executable, testable artifacts that compose like code, and suggests future IDEs will become 'integrated thought clarifiers.' He closes by calling for help in aligning agents at scale, noting OpenAI's new agent robustness team."
  },
  {
    "slug": "context-engineering-for-complex-codebases",
    "title": "No Vibes Allowed: Solving Hard Problems in Complex Codebases – Dex Horthy, HumanLayer",
    "url": "https://ai.engineer/talks/context-engineering-for-complex-codebases",
    "videoId": "rmvDxxNubIg",
    "event": null,
    "durationMs": 1231000,
    "speakers": [
      {
        "name": "Dex Horthy"
      }
    ],
    "topics": [
      {
        "slug": "coding-agents",
        "name": "Coding agents"
      }
    ],
    "summary": "Dex Horthy of HumanLayer argues that with deliberate context engineering — specifically a technique he calls \"frequent intentional compaction\" — today's AI coding agents can handle large brownfield codebases, not just greenfield projects. He presents a three-phase workflow (research, plan, implement) that keeps agents in the \"smart zone\" of the context window, avoiding the diminishing returns that set in around the 40% mark. Horthy demonstrates this approach solving a real problem in a 300k-line Rust codebase, shipping a week's worth of work in seven hours with code that passed expert review. He cautions against outsourcing thinking to AI, emphasising that plans must include actual code snippets for reliable execution and team alignment. The talk also addresses the cultural rift where senior engineers clean up slop from juniors' AI tools, calling for top-down adoption and team-wide workflow adaptation."
  },
  {
    "slug": "effective-ai-agents",
    "title": "How We Build Effective Agents: Barry Zhang, Anthropic",
    "url": "https://ai.engineer/talks/effective-ai-agents",
    "videoId": "D7_ipDqhtwk",
    "event": "summit-2025",
    "durationMs": 909000,
    "speakers": [
      {
        "name": "Barry Zhang"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Barry Zhang of Anthropic's Applied AI team argues that effective agents require simplicity, not complexity, and should be built only for tasks with high value and ambiguous problem spaces. He offers a checklist: ensure task complexity is high, value justifies token cost, critical capabilities are de-risked, and errors are easily discovered (e.g., coding with unit tests). Agents are just models using tools in a loop — environment, tools, and system prompt — and he advises iterating on these three components before optimizing. To improve agents, developers should think like them by narrowing their perspective to the agent's 10-20k token context window and even asking Claude to critique its own tools and trajectories. Zhang forecasts three open challenges: making agents budget-aware by enforcing time/token/money limits, enabling self-evolving tools via meta-tools, and building asynchronous multi-agent communication beyond synchronous turns."
  },
  {
    "slug": "pi-coding-agent",
    "title": "Building pi in a World of Slop — Mario Zechner",
    "url": "https://ai.engineer/talks/pi-coding-agent",
    "videoId": "RjfbvDXpFls",
    "event": null,
    "durationMs": 1105000,
    "speakers": [
      {
        "name": "Mario Zechner"
      }
    ],
    "topics": [
      {
        "slug": "coding-agents",
        "name": "Coding agents"
      }
    ],
    "summary": "Mario Zechner explains why he built pi, a self-modifying, extensible agent core, after growing frustrated with Claude Code and Open Code for their buggy features, uncontrolled context management, and zero extensibility. pi strips away complexity, giving the agent just four tools (read, write, edit, bash) and a minimal system prompt, letting it adapt to the user's workflow rather than vice versa. Extensions are TypeScript modules that hot-reload, enabling users to build custom tools, chat rooms, or even play Doom without forking pi. He then attacks clankers — agent-generated pull requests and issues — that are destroying open source, sharing his own tactics like auto-closing PRs with a human-voice request and de-prioritizing OpenClaw interactions. In Act 3, Zechner argues that agents compound errors with zero learning, producing enterprise-grade complexity in weeks, and that long context windows and agentic search are hacks. His prescription: scope agent tasks, modularize codebases, cap generated code review, and always read every line of critical code yourself."
  },
  {
    "slug": "production-ready-rag",
    "title": "Building Production-Ready RAG Applications: Jerry Liu",
    "url": "https://ai.engineer/talks/production-ready-rag",
    "videoId": "TRjq7t2Ms5I",
    "event": null,
    "durationMs": 1115000,
    "speakers": [
      {
        "name": "Jerry Liu"
      }
    ],
    "topics": [
      {
        "slug": "rag-and-knowledge",
        "name": "RAG and knowledge systems"
      }
    ],
    "summary": "Jerry Liu, CEO of LlamaIndex, explains how to productionize Retrieval Augmented Generation (RAG) systems by moving beyond naive implementations. He identifies key challenges: low retrieval precision causing hallucination, low recall from insufficient top-K, and lost-in-the-middle problems. Liu advocates starting with 'table stakes' improvements like tuning chunk sizes (showing optimal values per dataset), adding metadata filters (e.g., year=2021 for SEC 10Q queries), and hybrid search. More advanced techniques include 'small-to-big retrieval', embedding smaller chunks for precision then expanding windows for synthesis, and using reranking to improve recall. Finally, he explores agent architectures where each document becomes a tool for summarization or QA, and fine-tuning—generating synthetic query datasets from raw text to fine-tune embeddings, or distilling GPT-4's chain-of-thought into GPT-3.5 Turbo for better reasoning."
  },
  {
    "slug": "infinite-software-crisis",
    "title": "The Infinite Software Crisis – Jake Nations, Netflix",
    "url": "https://ai.engineer/talks/infinite-software-crisis",
    "videoId": "eIoohUmYpGI",
    "event": null,
    "durationMs": 1137000,
    "speakers": [
      {
        "name": "Jake Nations"
      }
    ],
    "topics": [
      {
        "slug": "coding-agents",
        "name": "Coding agents"
      }
    ],
    "summary": "Jake Nations, engineering lead at Netflix, argues that AI-generated code accelerates the software crisis by conflating easy with simple—producing tangled, incomprehensible systems. He traces the crisis from early software engineering debates to today's infinite code generation, citing Fred Brooks' 'No Silver Bullet' and Rich Hickey's definition of simple as 'one fold, no entanglement.' Nations presents a three-phase methodology—research (compressing 5 million tokens of code into a 2,000-word spec), planning (paint-by-numbers implementation steps), and implementation (using a manual migration seed)—to maintain human understanding. He warns that without this approach, engineers lose the ability to recognize dangerous complexity, and challenges listeners: will we still understand our own systems when AI writes most of the code?"
  },
  {
    "slug": "building-agents-with-mcp",
    "title": "Building Agents with Model Context Protocol - Full Workshop with Mahesh Murag of Anthropic",
    "url": "https://ai.engineer/talks/building-agents-with-mcp",
    "videoId": "kQmXtrmQ5Zg",
    "event": null,
    "durationMs": 6252000,
    "speakers": [
      {
        "name": "Mahesh Murag"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Mahesh Murag of Anthropic presents the Model Context Protocol (MCP) as an open standard that replaces fragmented integrations with a single protocol for connecting AI systems to data sources, enabling context-rich AI applications and agentic experiences. He explains MCP's philosophy, inspired by APIs and LSP, and its three interfaces: tools (model-controlled), resources (application-controlled), and prompts (user-controlled). Murag highlights adoption with over 1,100 community-built servers and official integrations from companies like Cloudflare and Stripe. He demonstrates building agents with MCP using the MCP-Agent framework, showing how agents can use tools dynamically and composably across hierarchical systems. Future plans include remote server support with OAuth 2.0, a centralized registry for discovery and verification, and enabling agents to self-evolve by dynamically finding new capabilities via registry search."
  },
  {
    "slug": "ai-developer-productivity",
    "title": "Does AI Actually Boost Developer Productivity? (100k Devs Study) - Yegor Denisov-Blanch, Stanford",
    "url": "https://ai.engineer/talks/ai-developer-productivity",
    "videoId": "tbDDYKRFjhk",
    "event": null,
    "durationMs": 1092000,
    "speakers": [
      {
        "name": "Yegor Denisov-Blanch"
      }
    ],
    "topics": [
      {
        "slug": "coding-agents",
        "name": "Coding agents"
      }
    ],
    "summary": "Yegor Denisov-Blanch from Stanford presents findings from a study of nearly 100,000 developers across 600+ companies, showing AI boosts developer productivity by an average of 15-20%, but the effect varies widely. The study measures functionality delivered, not commits, and reveals that AI introduces significant rework—bug fixes from AI-generated code. Productivity gains depend on task complexity, codebase maturity (Greenfield vs. Brownfield), language popularity, and codebase size: low-complexity Greenfield tasks see 30-40% gains, while high-complexity Brownfield tasks see only 0-10%. AI can even decrease productivity for low-popularity languages like COBOL or Haskell, and context window limitations reduce performance on larger codebases. The study provides data-driven guidance for when AI helps or hinders, with a matrix showing 20% gains for low-complexity common languages and minimal gains for complex niche ones."
  },
  {
    "slug": "ontologies-for-agentic-systems",
    "title": "Why Agentic Systems Need Ontologies — Frank Coyle, UC Berkeley",
    "url": "https://ai.engineer/talks/ontologies-for-agentic-systems",
    "videoId": "Sir59K8ZDPU",
    "event": null,
    "durationMs": 1278000,
    "speakers": [
      {
        "name": "Frank Coyle"
      }
    ],
    "topics": [
      {
        "slug": "rag-and-knowledge",
        "name": "RAG and knowledge systems"
      }
    ],
    "summary": "Frank Coyle argues that most agent failures, from brittle tools to fragile handoffs, stem from missing formal ontologies as logical guardrails. He proposes neurosymbolic AI: probabilistic reasoning inside, logic outside. An ontology is typed entities, relationships, and constraints expressed with RDFS and OWL, letting you specify a payment status must be one of three values, that a customer and support rep are different, or that an order can only be refunded once. Wrapping a Claude tool use loop with a validator—Pydantic at the door for types, ontology at the ledger for results—catches errors like a second refund on the same order, a payout sent to the support desk instead of the buyer, or an order status of 'probably shipped' that English instructions cannot reliably prevent."
  },
  {
    "slug": "12-factor-agents",
    "title": "12-Factor Agents: Patterns of reliable LLM applications — Dex Horthy, HumanLayer",
    "url": "https://ai.engineer/talks/12-factor-agents",
    "videoId": "8kMaTybvDUw",
    "event": null,
    "durationMs": 1026000,
    "speakers": [
      {
        "name": "Dex Horthy"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Dex Horthy, founder of HumanLayer, presents the 12-Factor Agents framework for building reliable LLM-powered applications, arguing that production-grade agents are primarily deterministic software with targeted LLM steps rather than fully autonomous loops. He distills patterns: own prompts and context windows, treat tools as JSON and code, use small focused agents with three to ten steps, contact humans via tool calls. Horthy emphasizes context engineering—LLMs are pure functions—and shows how to compact errors, unify state, and add pause/resume via APIs. He shares a DevOps agent that became a bash script, and advocates for outer-loop agents. The framework, which gained 4,000 GitHub stars in two months, treats agents as stateless reducers that meet users on any channel, with engineers controlling the inner loop of token and control flow."
  },
  {
    "slug": "graphrag",
    "title": "GraphRAG: The Marriage of Knowledge Graphs and RAG: Emil Eifrem",
    "url": "https://ai.engineer/talks/graphrag",
    "videoId": "knDDGYHnnSI",
    "event": "worldsfair-2024",
    "durationMs": 1155000,
    "speakers": [
      {
        "name": "Emil Eifrem"
      }
    ],
    "topics": [
      {
        "slug": "rag-and-knowledge",
        "name": "RAG and knowledge systems"
      }
    ],
    "summary": "Emil Eifrem, Neo4j co-founder and CEO, argues that GraphRAG—combining knowledge graphs with vector search—significantly improves RAG application accuracy, ease of development, and explainability. Citing studies, he reports accuracy gains of 3x (Data.org), 75-77% (LinkedIn), and Microsoft's finding that GraphRAG enables answering entirely new question types. He demonstrates Neo4j's Knowledge Graph Builder, which auto-generates graphs from PDFs, Wikipedia, and YouTube, making graph creation accessible. Eifrem frames this as the next evolution in search after PageRank and Google's knowledge graph, urging developers to adopt GraphRAG for richer context and better AI outcomes."
  },
  {
    "slug": "ai-agent-harnesses",
    "title": "Harnesses in AI: A Deep Dive — Tejas Kumar, IBM",
    "url": "https://ai.engineer/talks/ai-agent-harnesses",
    "videoId": "C_GG5g38vLU",
    "event": null,
    "durationMs": 1227000,
    "speakers": [
      {
        "name": "Tejas Kumar"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Tejas Kumar, AI Developer Advocate at IBM, defines an AI harness as everything around the model that grounds it in reality, contrasting it with an agent loop. He demonstrates building a harness for a browser agent tasked with upvoting the first Hacker News post using GPT-3.5 Turbo, emphasizing that the prompt remains unchanged. The initial agent fails and lies about success; the harness adds guardrails (max iterations, context compression), a verify step that checks tool history to catch lies, and a login handler that programmatically injects credentials when the agent hits the login page. The final agent reliably logs in and upvotes. Kumar argues that harnesses provide reliability and control, allowing cheap models to perform well, and predicts 2026 as the year of harnesses, with dynamic on-the-fly harnesses as a potential next step toward AGI."
  },
  {
    "slug": "evaluating-ai-agents",
    "title": "Building and evaluating AI Agents — Sayash Kapoor, AI Snake Oil",
    "url": "https://ai.engineer/talks/evaluating-ai-agents",
    "videoId": "d5EltXhbcfA",
    "event": "summit-2025",
    "durationMs": 1200000,
    "speakers": [
      {
        "name": "Sayash Kapoor"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Sayash Kapoor argues that current AI agents fall far short of their claimed performance due to flawed evaluation and a gap between capability and reliability. He cites failures like Do Not Pay (fined by FTC), LexisNexis (hallucinations in up to a third of cases), and Sakana AI (agent hacked reward functions, claiming 150x speedup that exceeded H100's theoretical max). Princeton's CoreBench shows best agents reproduce under 40% of papers. He emphasizes that agent benchmarks like SWE-bench mislead VC funding—Cognition's Devin succeeded on only 3 of 20 real-world tasks. Kapoor calls for cost-aware, multi-dimensional evaluation (e.g., Holistic Agent Leaderboard with Pareto frontiers) and a shift from capability to reliability engineering, drawing parallels to ENIAC's vacuum tube failures."
  },
  {
    "slug": "structured-llm-outputs-with-pydantic",
    "title": "Pydantic is all you need: Jason Liu",
    "url": "https://ai.engineer/talks/structured-llm-outputs-with-pydantic",
    "videoId": "yj-wSRJwrrc",
    "event": null,
    "durationMs": 1075000,
    "speakers": [
      {
        "name": "Jason Liu"
      }
    ],
    "topics": [
      {
        "slug": "rag-and-knowledge",
        "name": "RAG and knowledge systems"
      }
    ],
    "summary": "Jason Liu argues that Pydantic is the key to building reliable LLM applications by using structured prompting with OpenAI function calling. He introduces Instructor, a library that patches OpenAI's API to return Pydantic objects instead of raw JSON, ensuring type safety and validation. Liu demonstrates how Pydantic's 70 million downloads make it a trusted tool for defining data models with type hints, field validators, and even LLM-powered validators that catch errors like \"don't say mean things\" and retry via max retries. He shows concrete examples: decomposing user queries into structured search objects for RAG, generating a query plan DAG with parallel dependencies, extracting knowledge graphs for visualization, and verifying facts by requiring substring quotes from source text. Liu emphasizes that structuring prompts as code moves development from string manipulation to domain modeling, enabling cleaner, more maintainable systems that integrate easily with existing software."
  },
  {
    "slug": "production-multi-agent-architecture",
    "title": "The Multi-Agent Architecture That Actually Ships — Luke Alvoeiro, Factory",
    "url": "https://ai.engineer/talks/production-multi-agent-architecture",
    "videoId": "ow1we5PzK-o",
    "event": null,
    "durationMs": 1111000,
    "speakers": [
      {
        "name": "Luke Alvoeiro"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Luke Alvoeiro from Factory presents Missions, a multi-agent architecture that combines delegation, creator-verifier, broadcast, and negotiation into a three-role system of orchestrator, workers, and validators. The central claim is that human attention, not AI intelligence, is now the bottleneck in software engineering. Missions uses validation contracts written before any code to define correctness independently, structured handoffs to maintain context across 16-day runs, and serial execution with targeted internal parallelization to reduce errors. Strategic model selection per role—planning, implementation, validation—compounds advantages across model generations. Production data from building a Slack clone shows 60% of time and tokens spent on implementation, validation never succeeds on first go, and 50% of final code lines are tests with 90% coverage. The system is designed to improve with each model release by keeping orchestration logic in prompts and skills rather than hard-coded state machines."
  },
  {
    "slug": "harness-engineering",
    "title": "Harness Engineering: How to Build Software When Humans Steer, Agents Execute — Ryan Lopopolo, OpenAI",
    "url": "https://ai.engineer/talks/harness-engineering",
    "videoId": "am_oeAoUhew",
    "event": null,
    "durationMs": 2781000,
    "speakers": [
      {
        "name": "Ryan Lopopolo"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Ryan Lopopolo, a member of technical staff at OpenAI, argues that software engineers must shift to 'harness engineering' where agents produce all code and humans focus on delegation, system design, and guardrails. He describes how his team banned editors, using Codex with skills to manipulate apps, and reduced review friction by automating lint rules and review agents. Lopopolo explains that code is free but human time and model context are scarce, so teams should structure repositories for consistency and use 'garbage collection days' to eliminate recurring slop. He shares practical techniques like writing tests about source code (e.g., file length limits) and using review agents that inject prompts. His future vision involves agents autonomously executing long-horizon work based on token budgets and success metrics, freeing humans for higher-level activities."
  },
  {
    "slug": "production-rag-agents",
    "title": "RAG Agents in Prod: 10 Lessons We Learned — Douwe Kiela, creator of RAG",
    "url": "https://ai.engineer/talks/production-rag-agents",
    "videoId": "kPL-6-9MVyA",
    "event": "summit-2025",
    "durationMs": 1016000,
    "speakers": [
      {
        "name": "Douwe Kiela"
      }
    ],
    "topics": [
      {
        "slug": "rag-and-knowledge",
        "name": "RAG and knowledge systems"
      }
    ],
    "summary": "Douwe Kiela, CEO of Contextual AI and creator of RAG, shares 10 lessons from deploying enterprise RAG systems at scale. He argues that language models are only 20% of a larger system; success comes from focusing on systems, not models, and specializing over AGI to unlock domain expertise. Enterprise data is the real moat, but pilots are easy while production is hard—design for production from day one. Speed beats perfection: ship barely functional to real users early and iterate. Avoid boring engineering chores like chunking; instead, integrate AI into existing workflows to drive adoption. Accuracy is table stakes; handle inaccuracy with observability and attribution. Be ambitious: aim for transformative ROI, not low-hanging fruit like basic HR questions."
  },
  {
    "slug": "llm-pelican-benchmark",
    "title": "2025 in LLMs so far, illustrated by Pelicans on Bicycles — Simon Willison",
    "url": "https://ai.engineer/talks/llm-pelican-benchmark",
    "videoId": "YpY83-kA7Bo",
    "event": null,
    "durationMs": 1110000,
    "speakers": [
      {
        "name": "Simon Willison"
      }
    ],
    "topics": [],
    "summary": "Simon Willison reviews the past six months of LLM releases — including AWS Nova, Llama 3.3 70B, DeepSeek R1, Mistral Small 3, Claude 3.7 Sonnet, GPT 4.5, Gemini 2.5 Pro, GPT-4o, Llama 4, GPT 4.1, O3/O4 Mini, and Claude 4 — using his 'pelican on bicycle' SVG benchmark to argue that local models have become good enough to run GPT-4 class models on a laptop and that combining tools with reasoning is the most powerful technique in AI engineering, while noting risks like prompt injection and the 'lethal trifecta'. He tracks 30 significant model releases, highlighting that Mistral Small 3 (24B) matches Llama 3 70B's performance, which itself matched the 405B model, enabling local inference. DeepSeek's R1 caused a $500B+ Nvidia stock drop on January 27. GPT 4.1 Nano is the cheapest model yet at a fraction of a cent per pelican. He also examines bugs: ChatGPT's sycophantic 'shit-on-a-stick' incident and Claude 4's tendency to snitch to authorities when given ethical instructions and email tools. Willison concludes that while the pace is accelerating, control over context and security remain critical."
  },
  {
    "slug": "field-guide-to-fable",
    "title": "Field Guide to Fable — Thariq Shihipar, Anthropic",
    "url": "https://ai.engineer/talks/field-guide-to-fable",
    "videoId": "9fubhllmsBU",
    "event": null,
    "durationMs": 1168000,
    "speakers": [
      {
        "name": "Thariq Shihipar"
      }
    ],
    "topics": [],
    "summary": "Thariq Shihipar of Anthropic discusses Fable and argues models improve in \"spiky\" ways: a chat model fails to list Pokémon ending in \"aw\" (Croconaw and Dreadnaw), but Claude Code fetches and filters the list in seconds – a gap he calls \"capability overhang.\" He explains that to unlock Fable, Claude Code cut 80% of its system prompt because heavy instructions now constrain a more imaginative model, and the \"ask user question\" tool evolved from barely working under Opus 4 to generating embedded HTML questionnaires. He shares techniques like blind-spot passes and interviews to surface unknown unknowns, and reflects on the grief of moving from hand-coded programming to agentic workflows. Shihipar urges engineers to reject trade-offs – \"good, fast, cheap: pick three\" – and instead demand all three, citing a four-hour keynote deck built with Fable as proof that agents can deliver ambitious work faster."
  },
  {
    "slug": "rl-quantization-agents",
    "title": "[Full Workshop] Reinforcement Learning, Kernels, Reasoning, Quantization & Agents — Daniel Han",
    "url": "https://ai.engineer/talks/rl-quantization-agents",
    "videoId": "OkEGJ5G3foU",
    "event": null,
    "durationMs": 9748000,
    "speakers": [
      {
        "name": "Daniel Han"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Daniel Han of Unsloth presents a technical workshop covering reinforcement learning (RL), kernels, reasoning, quantization, and agents, arguing that RL with verifiable rewards (RLVR) is the key to unlocking LLM capabilities beyond supervised fine-tuning. He explains why open-source models plateaued after September 2024 until DeepSeek-R1 showed that RL can elicit reasoning, and breaks down PPO, GRPO, and the REINFORCE algorithm, emphasizing that GRPO removes the value model for efficiency. Han details how reward functions—not algorithms—are the hardest part, with examples like distance-based scoring for math. He demonstrates a free Colab notebook training a base model to reason, and shows that dynamic quantization can shrink models like DeepSeek-R1 from 730 GB to 140 GB with only ~1% accuracy loss, arguing that GPUs may stop getting faster after FP4 precision."
  },
  {
    "slug": "claude-agent-sdk-workshop",
    "title": "Claude Agent SDK [Full Workshop] — Thariq Shihipar, Anthropic",
    "url": "https://ai.engineer/talks/claude-agent-sdk-workshop",
    "videoId": "TqC1qOfiVcQ",
    "event": null,
    "durationMs": 6745000,
    "speakers": [
      {
        "name": "Thariq Shihipar"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Thariq Shihipar of Anthropic presents the Claude Agent SDK, arguing that Bash and file-system-based agents outperform traditional tool-only approaches for autonomous tasks. He defines agents as systems that build their own context and trajectory, contrasting with structured workflows. The SDK, built on Claude Code, emphasizes the Bash tool as the most powerful primitive for composability and code generation, enabling non-coding tasks like data analysis. He demonstrates live-coding a Pokémon team advisor that dynamically fetches API data via scripts, and explains security through a 'Swiss cheese defense' of model alignment, AST parsing, and sandboxing. Shihipar also covers skills for progressive context disclosure, sub-agents for parallel work, and hooks for deterministic verification, stressing that agent building is an art of reading transcripts and iterating on context engineering."
  },
  {
    "slug": "generative-ui-for-mcp",
    "title": "Beyond Components: Designing Generative UI for MCP Apps — Ruben Casas, Postman",
    "url": "https://ai.engineer/talks/generative-ui-for-mcp",
    "videoId": "hCMrEfPG2Yg",
    "event": null,
    "durationMs": 1018000,
    "speakers": [
      {
        "name": "Ruben Casas"
      }
    ],
    "topics": [],
    "summary": "Ruben Casas from Postman argues that AI models can now write better frontend code than he can—his prompt to rewrite his blog produced a search box with blur animation and accessibility out of the box—yet most agent UIs still invoke static prebuilt components. He presents three levels of UI generation: static components (AG UI, Goose) passing props to predefined React elements; declarative UI where the model generates JSON or YAML for a rendering engine (e.g., Vercel's JSON Render), which he deems the current ideal balance; and fully generative UI where the model writes HTML, CSS, and JavaScript on demand, as in his weather agent that does so in one tool call. The key obstacle is trust, necessitating sandboxing, and MCP apps with their double iframe default are the best delivery mechanism. He likens today to early TV—radio shows with cameras—and predicts the future lies beyond components in collaborative human-agent interfaces on shared canvases, as seen with the Skeletro MCP app."
  },
  {
    "slug": "state-of-the-claw",
    "title": "State of the Claw — Peter Steinberger",
    "url": "https://ai.engineer/talks/state-of-the-claw",
    "videoId": "zgNvts_2TUE",
    "event": null,
    "durationMs": 2652000,
    "speakers": [
      {
        "name": "Peter Steinberger"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Peter Steinberger, creator of OpenClaw, presents a five-month update on the world's fastest-growing open-source project. He details the project's staggering growth—30,000 commits, nearly 2,000 contributors—and the immense security burden: 1,142 advisories (16.6 per day) with 99 critical, often AI-generated slop that demands human vetting. Steinberger refutes media fearmongering, citing how researchers ignore security docs to fabricate scary scenarios. He clarifies OpenAI did not buy OpenClaw; he joined the company while establishing the OpenClaw Foundation to remain vendor-neutral. He emphasizes the importance of local models for data sovereignty and describes his coding workflow of running 5-6 agent sessions simultaneously, iterating on taste and personality. Future visions include ubiquitous agents, 'Dreaming' for memory reconciliation, and modular plugins. For engineers, he champions taste, system design, and learning to say no."
  },
  {
    "slug": "future-of-mcp",
    "title": "The Future of MCP — David Soria Parra, Anthropic",
    "url": "https://ai.engineer/talks/future-of-mcp",
    "videoId": "v3Fr2JR47KA",
    "event": null,
    "durationMs": 1126000,
    "speakers": [
      {
        "name": "David Soria Parra"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "David Soria Parra from Anthropic argues that MCP (Model Context Protocol) is the key to connecting agents to tools and data in production, with 110 million monthly downloads—outpacing React's growth at the same stage. He lays out a 2026 connectivity stack combining Skills, MCP, and CLI/Computer Use, each suited for different needs, and emphasizes that best agents will use all three seamlessly. To improve client harnesses, he introduces Progressive Discovery—deferring tool loading via Tool Search to reduce context usage—and Programmatic Tool Calling, where models write scripts to compose tool outputs efficiently. Upcoming MCP protocol improvements include stateless transport (with Google) for easier scaling, async agent-to-agent tasks, enterprise features like Cross App Access and Server Discovery via well-known URLs, and a Skills-over-MCP extension for shipping usage instructions with servers. He calls for community feedback on these directions."
  },
  {
    "slug": "future-of-knowledge-assistants",
    "title": "The Future of Knowledge Assistants: Jerry Liu",
    "url": "https://ai.engineer/talks/future-of-knowledge-assistants",
    "videoId": "zeAyuLc_f3Q",
    "event": "worldsfair-2024",
    "durationMs": 1015000,
    "speakers": [
      {
        "name": "Jerry Liu"
      }
    ],
    "topics": [
      {
        "slug": "rag-and-knowledge",
        "name": "RAG and knowledge systems"
      }
    ],
    "summary": "Jerry Liu, CEO of LlamaIndex, explains how to move beyond simple RAG to build production-grade knowledge assistants. He details three steps: advanced data processing with LlamaParse for accurate PDF parsing, single-agent flows with query planning and tool use, and multi-agent task solvers via the newly announced Llama Agents framework. Llama Agents treats each agent as a deployable microservice that communicates through a central API, enabling specialization, parallelism, and easier production deployment. Jerry also highlights that naive RAG is insufficient for complex queries, and that good data quality—like proper parsing of tables and charts—is essential to reduce hallucinations."
  },
  {
    "slug": "pruning-agent-skills",
    "title": "How I deleted 95% of my agent skills and got better results — Nick Nisi, WorkOS",
    "url": "https://ai.engineer/talks/pruning-agent-skills",
    "videoId": "vy7o1g2iHY8",
    "event": null,
    "durationMs": 1063000,
    "speakers": [
      {
        "name": "Nick Nisi"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Nick Nisi, DX engineer at WorkOS, argues that AI agents should be forced to prove their work with code rather than trusted with prompts. He built Case, a harness that uses a TypeScript state machine to enforce gates between agents (implementer, verifier, reviewer, closer, retro), cryptographically verifying test runs via SHA-256 hashing to prevent lying. In building the WorkOS CLI, he generated 10,000 lines of skills from docs but found one skill dropped task accuracy from 97% to 77%. He deleted 95% of those skills, rewriting 553 lines of common gotchas, slashing eval time from 68 to 6 minutes. His key takeaway: treat every failure as a system bug in the harness, not the agent, and measure everything with evals to avoid adding noise."
  },
  {
    "slug": "building-great-agent-skills",
    "title": "Building Great Agent Skills: The Missing Manual",
    "url": "https://ai.engineer/talks/building-great-agent-skills",
    "videoId": "UNzCG3lw6O0",
    "event": "worldsfair-2026-online-track",
    "durationMs": 1243000,
    "speakers": [
      {
        "name": "Matt Pocock"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-agents",
        "name": "Coding agents"
      }
    ],
    "summary": "Matt Pocock presents a 'skill checklist' framework to help developers escape 'skill hell' by systematically evaluating and improving agent skills, covering trigger, structure, steering, and pruning. He explains the trade-offs between user-invoked and model-invoked skills: model-invoked increase context load and unpredictability, user-invoked increase pilot cognitive load. Structure divides skills into steps and reference, with branching material hidden behind context pointers to minimize skill.md. Steering uses 'leading words' like 'vertical slice' that appear in reasoning traces; to boost leg work, he splits skills so agents only see one step at a time. Pruning removes duplication, sediment, and no-ops that don't affect behavior. The framework is in the 'writing-great-skills' skill in his repository."
  },
  {
    "slug": "1-ai-guardrails-the-unreasonable-effectiveness-of-finetuned-modernberts-diego-carpentero",
    "title": "$1 AI Guardrails: The Unreasonable Effectiveness of Finetuned ModernBERTs – Diego Carpentero",
    "url": "https://ai.engineer/talks/1-ai-guardrails-the-unreasonable-effectiveness-of-finetuned-modernberts-diego-carpentero",
    "videoId": "YZHPEkfy2kc",
    "event": "AI Engineer Europe 2026",
    "durationMs": 2633000,
    "speakers": [
      {
        "slug": "diego-carpentero",
        "name": "Diego Carpentero"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Diego Carpentero examines prompt and context injection, retrieval and Model Context Protocol attack surfaces, and agentic escalation, using Bing Chat and Wikipedia examples to explain why LLMs struggle to separate trusted instructions from untrusted data. He proposes fine-tuning ModernBERT into a low-latency, self-hosted safety discriminator for under a dollar, explains alternating attention, rotary positional encoding, FlashAttention, CLS-token representations, unpadding, and sequence packing, and walks through dataset preparation and practical safety checks."
  },
  {
    "slug": "120k-players-in-a-week-lessons-from-the-first-viral-clip-app",
    "title": "120k players in a week: Lessons from the first viral CLIP app",
    "url": "https://ai.engineer/talks/120k-players-in-a-week-lessons-from-the-first-viral-clip-app",
    "videoId": "OimPoLxioYg",
    "event": "AI Engineer Summit 2023",
    "durationMs": 959000,
    "speakers": [
      {
        "slug": "redacted-username-nelson",
        "name": "[REDACTED:username] Nelson"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Roboflow cofounder [REDACTED:username] Nelson explains how Paint.wtf attracted 120,000 players in its first week by combining GPT-3-generated drawing prompts, a browser-based canvas, CLIP image-text embeddings, cosine-similarity scoring, and a Supabase-powered leaderboard. He demonstrates a compact Python implementation using OpenCV, Roboflow Inference, and Roboflow Universe models, then discusses open-set visual understanding, model limitations, and penalizing handwritten text to prevent players from gaming CLIP-based scores."
  },
  {
    "slug": "20-days-of-compute-vs-7-hours-rethinking-what-state-of-the-art-means-bertrand-charpentier-pruna",
    "title": "20 days of compute vs 7 hours: rethinking what state-of-the-art means — Bertrand Charpentier, Pruna AI",
    "url": "https://ai.engineer/talks/20-days-of-compute-vs-7-hours-rethinking-what-state-of-the-art-means-bertrand-charpentier-pruna",
    "videoId": "hqHC6Z_lXyo",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1176000,
    "speakers": [
      {
        "slug": "bertrand-charpentier",
        "name": "Bertrand Charpentier",
        "organization": "Pruna AI"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Pruna AI cofounder and chief scientist Bertrand Charpentier argues that state-of-the-art image models cannot be identified from a single leaderboard or quality metric. Comparing Design Arena, Arena, and Artificial Analysis, he shows how rankings vary by benchmark and use case, demonstrates limitations of human inspection and CLIP scores, and highlights the compute and energy costs of large-scale evaluations. He advocates assessing specialized, compressed models on a Pareto front that balances task-specific quality against latency, cost, and efficiency, followed by audience questions about image-model and video-model compression."
  },
  {
    "slug": "200-million-patient-interactions-later-what-the-generic-voice-stack-misses",
    "title": "200 Million Patient Interactions Later: What the Generic Voice Stack Misses",
    "url": "https://ai.engineer/talks/200-million-patient-interactions-later-what-the-generic-voice-stack-misses",
    "videoId": "AN65uc645mE",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1240000,
    "speakers": [
      {
        "slug": "vivek-muppalla",
        "name": "Vivek Muppalla",
        "organization": "Hippocratic AI"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Hippocratic AI engineering executive Vivek Muppalla explains how healthcare voice agents can expand proactive patient outreach without replacing clinicians. Drawing on more than 200 million reported patient interactions, he demonstrates clinical conversations, medication-sensitive guidance, and nurse escalation, then describes Polaris: a central conversational model supported by 30 specialist models and tool-call verifiers. He also outlines healthcare-specific evaluation, context-aware decoder-only audio processing, a clinically fine-tuned Whisper large-v3-turbo component, and the need to balance low latency, safety, quality, and empathy."
  },
  {
    "slug": "2025-is-the-year-of-evals-just-like-2024-and-2023-and",
    "title": "2025 is the Year of Evals!  Just like 2024, and 2023, and …",
    "url": "https://ai.engineer/talks/2025-is-the-year-of-evals-just-like-2024-and-2023-and",
    "videoId": "CQGuvf6gSrM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1154000,
    "speakers": [
      {
        "slug": "john-dickerson",
        "name": "John Dickerson",
        "organization": "Mozilla.ai"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Mozilla.ai CEO and former Arthur AI co-founder John Dickerson argues that AI evaluation and monitoring become commercially urgent as generative AI and autonomous agents reach production. He connects executive attention, enterprise budgets, measurable business KPIs, security, and quantitative risk to the shift from conventional ML observability toward agentic and multi-agent monitoring. Audience questions address domain-specific evaluation and the need to validate LLM-based judges against bias."
  },
  {
    "slug": "2026-the-year-the-ide-died",
    "title": "2026: The Year the IDE Died",
    "url": "https://ai.engineer/talks/2026-the-year-the-ide-died",
    "videoId": "7Dtu2bilcFs",
    "event": "AI Engineer Code 2025",
    "durationMs": 1499000,
    "speakers": [
      {
        "slug": "steve-yegge",
        "name": "Steve Yegge",
        "organization": "Sourcegraph/Amp"
      },
      {
        "slug": "gene-kim",
        "name": "Gene Kim",
        "organization": "IT Revolution"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Steve Yegge and Gene Kim argue that traditional IDEs will give way to AI-native development environments and increasingly orchestrated coding agents. Yegge compares current tools such as Claude Code to handheld power tools and future systems to CNC machines, discusses developer resistance, context-window limits, task decomposition, and products including Codex and Replit. Kim connects vibe coding to DevOps, organizational change, changing team composition, and leadership workshops that help participants build software with AI."
  },
  {
    "slug": "3-ingredients-for-building-reliable-enterprise-agents",
    "title": "3 ingredients for building reliable enterprise agents",
    "url": "https://ai.engineer/talks/3-ingredients-for-building-reliable-enterprise-agents",
    "videoId": "kTnfJszFxCg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1255000,
    "speakers": [
      {
        "slug": "harrison-chase",
        "name": "Harrison Chase",
        "organization": "LangChain"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "LangChain co-founder Harrison Chase explains how enterprise agents move beyond prototypes by targeting valuable, substantial work; combining agent autonomy with deterministic workflows through LangGraph; and making behavior observable and evaluable with LangSmith. He also highlights reversible actions, human correction, and interface patterns that reduce perceived risk, using legal research, coding agents, and Replit Agent as examples, before an audience question about measurability and reversibility."
  },
  {
    "slug": "360brew-llm-based-personalized-ranking-and-recommendation-hamed-firooz-and-maziar-sanjabi-linked",
    "title": "360Brew: LLM-based Personalized Ranking and Recommendation — Hamed Firooz and Maziar Sanjabi, LinkedIn AI",
    "url": "https://ai.engineer/talks/360brew-llm-based-personalized-ranking-and-recommendation-hamed-firooz-and-maziar-sanjabi-linked",
    "videoId": "U0S6CfzAY5c",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1320000,
    "speakers": [
      {
        "slug": "hamed-firooz",
        "name": "Hamed Firooz",
        "organization": "LinkedIn"
      },
      {
        "slug": "maziar-sanjabi",
        "name": "Maziar Sanjabi",
        "organization": "LinkedIn"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Hamed Firooz and Maziar Sanjabi describe LinkedIn's 360Brew, a unified foundation-model approach to personalized ranking and recommendation. They explain Brew-XL training and alignment, its 150-billion-parameter scale, experiments with Mixtral and longer user-history contexts, and deployment techniques for high-throughput, low-latency serving, including pruning and quantization. Audience questions address generalization, NLWeb, multi-item scoring, user history, and experimentation infrastructure."
  },
  {
    "slug": "500-people-vibe-coded-for-30-days-i-was-one-of-them",
    "title": "500 people vibe-coded for 30 days. I was one of them.",
    "url": "https://ai.engineer/talks/500-people-vibe-coded-for-30-days-i-was-one-of-them",
    "videoId": "UcYoMg-8-L8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1069000,
    "speakers": [
      {
        "slug": "sanja-grbic",
        "name": "Sanja Grbic",
        "organization": "Automattic"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Automattic product designer Sanja Grbic describes Radical Speed Month, a 30-day experiment in which roughly 500 employees initiated 794 projects through autonomous small-team work. She explains how role-specific AI training, accessible development environments, an internal MCP knowledge server, and Claude Code helped her contribute code and build three projects, including a design-system tracker and a WordPress.com/Jetpack-connected AI prototype. Her central lesson is that organizational speed depends on employee agency, cross-functional collaboration, and engineers enabling colleagues beyond traditional role boundaries."
  },
  {
    "slug": "6-things-to-know-about-aie-world-s-fair-2026",
    "title": "6 Things to Know about AIE World's Fair 2026",
    "url": "https://ai.engineer/talks/6-things-to-know-about-aie-world-s-fair-2026",
    "videoId": "0S8xe9ftGTM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1070000,
    "speakers": [
      {
        "slug": "swyx",
        "name": "Shawn \"swyx\" Wang",
        "organization": "AI Engineer"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "AI Engineer co-founder swyx previews the 2026 World's Fair, emphasizing its expanded technical program, Autoresearch and model-infrastructure topics, Moscone West expo and hallway networking, research-industry connections, leadership and Token Billionaire programming, industry-specific AI verticals, and surrounding community events."
  },
  {
    "slug": "7-habits-of-highly-effective-generative-ai-evaluations",
    "title": "7 Habits of Highly Effective Generative AI Evaluations",
    "url": "https://ai.engineer/talks/7-habits-of-highly-effective-generative-ai-evaluations",
    "videoId": "wHhlvcQgi9M",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1539000,
    "speakers": [
      {
        "slug": "justin-muller",
        "name": "Justin Muller",
        "organization": "Amazon Web Services (AWS)"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "AWS principal applied AI architect Justin Muller explains why rigorous evaluation frameworks are essential for moving generative AI workloads beyond proofs of concept. He argues evaluations should primarily expose actionable failures rather than simply produce scores, demonstrates how plausible outputs can conceal faulty reasoning and arithmetic, recommends averaging across sufficient test cases, and cautions against using generative AI to create gold-standard datasets before analyzing results by category."
  },
  {
    "slug": "a-genius-with-amnesia",
    "title": "A Genius With Amnesia",
    "url": "https://ai.engineer/talks/a-genius-with-amnesia",
    "videoId": "jVjt-2g8NMY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1200000,
    "speakers": [
      {
        "slug": "victor-savkin",
        "name": "Victor Savkin",
        "organization": "Nx"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Victor Savkin argues that capable coding agents remain constrained by fragmented repository visibility and the inability to remember previous sessions, forcing developers to repeatedly explain changes and coordinate downstream fixes. He presents Nx-built Polygraph, an agent-agnostic meta-harness that maps dependencies across repositories, coordinates changes and continuous integration, and preserves shared organizational context across agent sessions."
  },
  {
    "slug": "a-piece-of-pi-embedding-the-openclaw-coding-agent-in-your-product",
    "title": "A Piece of PI – Embedding The OpenClaw Coding Agent In Your Product",
    "url": "https://ai.engineer/talks/a-piece-of-pi-embedding-the-openclaw-coding-agent-in-your-product",
    "videoId": "vAIDdLKB6-w",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1242000,
    "speakers": [
      {
        "slug": "matthias-luebken",
        "name": "Matthias Luebken",
        "organization": "TAVON.ai"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "TAVON.ai founder Matthias Luebken explains how to embed Pi, the minimal coding-agent toolkit used by OpenClaw, into business products. He frames agents as language models repeatedly calling tools, advocates simple agent-friendly interfaces and composable command-line utilities, and demonstrates CRM lead qualification, TypeScript extensions, interactive controls, and inbox-driven sales workflows."
  },
  {
    "slug": "a-practical-guide-to-efficient-ai",
    "title": "A Practical Guide to Efficient AI",
    "url": "https://ai.engineer/talks/a-practical-guide-to-efficient-ai",
    "videoId": "Yyg_BoeB2LU",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1065000,
    "speakers": [
      {
        "slug": "shelby-heinecke",
        "name": "Shelby Heinecke",
        "organization": "Salesforce"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Salesforce AI research leader Shelby Heinecke explains how efficiency helps teams move AI systems from demonstrations into enterprise production and onto resource-constrained devices. She discusses efficient architectures, smaller language models such as the 3.8-billion-parameter Phi-3-mini, post-training quantization from 32-bit to 4-bit weights, and MobileAIBench for assessing on-device language and multimodal models."
  },
  {
    "slug": "a-practitioner-s-guide-to-graphs-tim-ainge-good-collective",
    "title": "A Practitioner's Guide to Graphs - Tim Ainge, Good Collective",
    "url": "https://ai.engineer/talks/a-practitioner-s-guide-to-graphs-tim-ainge-good-collective",
    "videoId": "3ySF0I5iE_0",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 858000,
    "speakers": [
      {
        "slug": "tim-ainge",
        "name": "Tim Ainge",
        "organization": "Good Collective"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Tim Ainge of Good Collective explains when graph-native techniques improve AI applications without treating GraphRAG as a universal solution. He discusses structured graph construction and entity matching, compares Cypher with SQL for graph traversal, and demonstrates personalized PageRank through Pinterest Pixie and HippoRAG. Additional examples cover legal citation discovery, code-context retrieval and decorator-pattern detection through subgraph matching in a .NET eShop codebase; he reports a 40% reduction in code-search tool calls in one evaluation."
  },
  {
    "slug": "a-song-of-types-and-agents",
    "title": "A Song of Types and Agents",
    "url": "https://ai.engineer/talks/a-song-of-types-and-agents",
    "videoId": "UlFB6efYN5Q",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 856000,
    "speakers": [
      {
        "slug": "roberto-stagi",
        "name": "Roberto Stagi",
        "organization": "Ratel"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Ratel co-founder Roberto Stagi argues that AI is moving from Python-centered model training and research into TypeScript-centered application and agent development. Drawing on GitHub's 2024 and 2025 language rankings, coding-agent adoption, Anthropic's acquisition of Bun, npm's ecosystem, and end-to-end type safety, he makes the case for building production AI agents within a unified TypeScript application stack while acknowledging Python's continuing role in training and GPU serving."
  },
  {
    "slug": "a-taxonomy-for-next-generation-reasoning-models",
    "title": "A Taxonomy for Next-Generation Reasoning Models",
    "url": "https://ai.engineer/talks/a-taxonomy-for-next-generation-reasoning-models",
    "videoId": "jQcsVk0KWiQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1161000,
    "speakers": [
      {
        "slug": "nathan-lambert",
        "name": "Nathan Lambert",
        "organization": "Allen Institute for AI (Ai2); Interconnects.ai"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Nathan Lambert outlines a four-part taxonomy for next-generation reasoning models—skills, calibration, strategy, and abstraction—and argues that progress beyond benchmark performance requires deliberately training planning and long-horizon agent capabilities. Drawing on examples including o3, Deep Research, Claude Code, and Codex, he connects reinforcement learning with verifiable rewards and inference-time scaling to emerging agent applications, while emphasizing token costs, supervised fine-tuning, reinforcement learning for larger tasks, and the possibility that post-training compute approaches pre-training compute."
  },
  {
    "slug": "a-year-of-gemini-progress-what-comes-next",
    "title": "A year of Gemini progress + what comes next",
    "url": "https://ai.engineer/talks/a-year-of-gemini-progress-what-comes-next",
    "videoId": "wE1ZCmCLP5g",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 718000,
    "speakers": [
      {
        "slug": "logan-kilpatrick",
        "name": "Logan Kilpatrick",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Google DeepMind's Logan Kilpatrick introduces an updated Gemini 2.5 Pro model, reviews a year of Gemini progress, and explains how specialized research including AlphaProof and AlphaGeometry informs Google's mainline models. He describes the Gemini app and developer API, then discusses proactive assistants, native audio, Veo video generation, model reasoning, embeddings for RAG, and a prospective Deep Research API."
  },
  {
    "slug": "a2a-mcp-automating-business-processes-with-llms",
    "title": "A2A & MCP: Automating Business Processes with LLMs",
    "url": "https://ai.engineer/talks/a2a-mcp-automating-business-processes-with-llms",
    "videoId": "wXVvfFMTyzY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 4994000,
    "speakers": [
      {
        "slug": "damien-murphy",
        "name": "Damien Murphy",
        "organization": "Bench Computing"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Bench Computing's Damien Murphy leads a hands-on workshop combining Google's Agent2Agent protocol with Model Context Protocol to automate business processes using cooperating AI agents. The session covers task delegation, integrations with Slack, GitHub, Salesforce, and Zapier, host-agent orchestration, OAuth and user-specific permissions, implementation-language tradeoffs, and audience questions about MCP interfaces and testing against real external services."
  },
  {
    "slug": "accelerate-your-ai-journey-with-azure-ai-model-catalog",
    "title": "Accelerate your AI journey with Azure AI model catalog",
    "url": "https://ai.engineer/talks/accelerate-your-ai-journey-with-azure-ai-model-catalog",
    "videoId": "gL9kfxt6uo0",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1394000,
    "speakers": [
      {
        "slug": "sharmila-chokalingam",
        "name": "Sharmila Chokalingam",
        "organization": "Microsoft"
      },
      {
        "slug": "shubhi",
        "name": "Shubhi"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Sharmila Chokalingam and a second presenter identifying herself as Shubhi demonstrate how Azure AI Model Catalog supports foundation-model selection, standardized inference, retrieval-augmented generation, deployment, and enterprise safeguards. They survey GPT-4o, Phi-3, Mistral, Llama, Cohere, and Jais; demonstrate Cohere Command R, LangChain/LiteLLM integrations, Azure Marketplace billing, benchmarks, playgrounds, and prompt flow; and discuss privacy, security, and [REDACTED:url]ai."
  },
  {
    "slug": "accelerating-ai-on-edge-chintan-parikh-and-weiyi-wang-google-deepmind",
    "title": "Accelerating AI on Edge — Chintan Parikh and Weiyi Wang, Google DeepMind",
    "url": "https://ai.engineer/talks/accelerating-ai-on-edge-chintan-parikh-and-weiyi-wang-google-deepmind",
    "videoId": "Lm8BLHkxiAo",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1438000,
    "speakers": [
      {
        "slug": "chintan-parikh",
        "name": "Chintan Parikh",
        "organization": "Google"
      },
      {
        "slug": "weiyi-wang",
        "name": "Weiyi Wang",
        "organization": "Google"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "Google LiteRT product manager Chintan Parikh introduces Google AI Edge and Gemma 4 edge models, describing the shift toward reasoning-capable on-device agents. The presentation demonstrates local agent and robotics use cases, explains LiteRT's TensorFlow Lite foundation and cross-platform Android, iOS, and IoT deployment, and covers AI Edge Portal benchmarking plus Qualcomm and MediaTek hardware integrations. Colleague Weiyi Wang is introduced for the question-and-answer portion, which includes questions about cameras, Raspberry Pi, and Orin devices."
  },
  {
    "slug": "accelerating-mixture-of-experts-training-with-rail-optimized-infiniband-networking-in-crusoe-clo",
    "title": "Accelerating Mixture of Experts Training With Rail-Optimized InfiniBand Networking in Crusoe Cloud",
    "url": "https://ai.engineer/talks/accelerating-mixture-of-experts-training-with-rail-optimized-infiniband-networking-in-crusoe-clo",
    "videoId": "tQTB4MU_z8w",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1065000,
    "speakers": [
      {
        "slug": "ievgen-vakulenko",
        "name": "Ievgen Vakulenko",
        "organization": "Crusoe"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Crusoe product manager Ievgen Vakulenko explains how Crusoe Cloud supports distributed AI training with GPU-equipped virtual machines, CPU compute, local NVMe storage, conventional VPC networking, and a separate rail-optimized InfiniBand fabric for high-bandwidth, low-latency GPU communication. He also discusses energy-conscious data-center infrastructure, developer usability, and customer workloads involving Together AI and Boson AI."
  },
  {
    "slug": "active-graph-agent-runtime-babyagi-4",
    "title": "Active Graph Agent Runtime (BabyAGI 4)",
    "url": "https://ai.engineer/talks/active-graph-agent-runtime-babyagi-4",
    "videoId": "khVX_BUnEwU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1054000,
    "speakers": [
      {
        "slug": "yohei-nakajima",
        "name": "Yohei Nakajima",
        "organization": "Untapped Capital"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Yohei Nakajima of Untapped Capital presents ActiveGraph, an experimental event-sourced graph runtime developed from his BabyAGI and graph-agent work. Instead of centering agents on LLM conversations, ActiveGraph treats a typed, immutable event log as the source of truth and projects shared graph state that reactive behaviors update, enabling auditability, replay, rollback, and forking. Nakajima explains policy-gated changes, modular packs, graph-based retrieval, blackboard-style coordination, and early self-improving agent experiments."
  },
  {
    "slug": "adaption-labs-gradient-free-continual-learning",
    "title": "Adaption Labs — Gradient-Free Continual Learning",
    "url": "https://ai.engineer/talks/adaption-labs-gradient-free-continual-learning",
    "videoId": "XEd_SRVHBgU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1251000,
    "speakers": [
      {
        "slug": "sara-hooker",
        "name": "Sara Hooker",
        "organization": "Adaption"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Sara Hooker examines how concentrated compute resources and narrow research career paths restrict participation in frontier AI. She presents Adaption's AutoScientist as an approach to automating model training and combining adaptive data, domain knowledge, and self-improvement, then argues that efficient customization, alternatives to brute-force pre-training, and competitive smaller models can broaden access. Audience questions address democratization, GPU purchasing dynamics, and open-source risks."
  },
  {
    "slug": "agent-evals-finally-with-the-map",
    "title": "Agent Evals: Finally, With The Map",
    "url": "https://ai.engineer/talks/agent-evals-finally-with-the-map",
    "videoId": "y2Drx0SDZLo",
    "event": "AI Engineer Summit 2025",
    "durationMs": 811000,
    "speakers": [
      {
        "slug": "ari-heljakka",
        "name": "Ari Heljakka",
        "organization": "Root Signals"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Root Signals presenter Ari Heljakka introduces a structured map for evaluating AI agents, separating semantic quality from behavioral performance. He discusses single-turn and multi-turn evaluation, RAG grounding and faithfulness, JSON tool-use validation, and API-specific metrics. He concludes that reliable agent evaluation requires EvalOps: optimizing both the operative LLM workflow and the separate judgment workflow used to evaluate it."
  },
  {
    "slug": "agent-output-is-not-ux-rendering-layer-your-llm-pipeline-is-missing-bala-ramdoss-amazon-lens",
    "title": "Agent Output Is Not UX: Rendering Layer Your LLM Pipeline Is Missing - Bala Ramdoss, Amazon Lens",
    "url": "https://ai.engineer/talks/agent-output-is-not-ux-rendering-layer-your-llm-pipeline-is-missing-bala-ramdoss-amazon-lens",
    "videoId": "maTp79FD9gI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 853000,
    "speakers": [
      {
        "slug": "bala-ramdoss",
        "name": "Bala Ramdoss",
        "organization": "Amazon"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Amazon engineer Bala Ramdoss argues that useful agentic products require a dedicated delivery layer that transforms model output into actionable, trustworthy interfaces. Drawing on Amazon Lens and Lens Live, he explains Google's A2UI and the spectrum from controlled components to open-ended generated interfaces, then outlines three production patterns: version-aware rendering contracts, progressive structured-UI streaming, and Backend-for-Frontend services that handle safe fallbacks and platform-specific rendering."
  },
  {
    "slug": "agent-reinforcement-fine-tuning",
    "title": "Agent Reinforcement Fine Tuning",
    "url": "https://ai.engineer/talks/agent-reinforcement-fine-tuning",
    "videoId": "p1CmPZ2j6Lk",
    "event": "AI Engineer Code 2025",
    "durationMs": 1015000,
    "speakers": [
      {
        "slug": "will-hang",
        "name": "Will Hang",
        "organization": "OpenAI"
      },
      {
        "slug": "cathy-zhou",
        "name": "Cathy Zhou",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "OpenAI's Will Hang and Cathy Zhou explain Agent RFT, which improves tool-using reasoning agents by training against custom reward signals and graders. They discuss public tool endpoints, rollout-level UUID tracking, domain adaptation, sample efficiency, inference-budget penalties, and production-representative evaluations. Coding-agent examples include OpenAI Codex, Cognition's Devin with isolated virtual-machine rollouts, and Cosine, followed by guidance on exploration and continuous rewards."
  },
  {
    "slug": "agentcraft-putting-the-orc-in-agent-orchestration",
    "title": "AgentCraft: Putting the Orc in Agent Orchestration",
    "url": "https://ai.engineer/talks/agentcraft-putting-the-orc-in-agent-orchestration",
    "videoId": "kR64LOqBBCU",
    "event": "AI Engineer Europe 2026",
    "durationMs": 678000,
    "speakers": [
      {
        "slug": "ido-salomon",
        "name": "Ido Salomon"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Ido Salomon demonstrates AgentCraft, an RTS-inspired interface designed to reduce the human bottleneck in orchestrating many coding agents. The demo covers support for existing agent tools, visual file-system and agent-activity tracking, fast approval workflows, bundled change review, and shared workspaces where teammates can collaborate across their respective agents."
  },
  {
    "slug": "agentic-development-security",
    "title": "Agentic Development Security",
    "url": "https://ai.engineer/talks/agentic-development-security",
    "videoId": "cgimkNGNjvU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1653000,
    "speakers": [
      {
        "slug": "ezra-tanzer",
        "name": "Ezra Tanzer",
        "organization": "Snyk"
      },
      {
        "slug": "dan-arpino",
        "name": "Dan Arpino",
        "organization": "Snyk"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Snyk product director Ezra Tanzer examines the security risks created when autonomous coding agents connect to external tools through MCP, including excessive privileges, vulnerable integrations, and destructive agent actions. He describes scanning and remediating AI-generated code and the broader need to inspect agent tooling and behavior. Snyk engineer Dan Arpino demonstrates an in-development local Electron security assistant that monitors agent sessions, exposes activity, applies guardrails, and can automatically address vulnerable tools. The discussion closes by contrasting security teams’ desire for strict controls with developers’ sensitivity to false positives."
  },
  {
    "slug": "agentic-engineering-working-with-ai-not-just-using-it-brendan-o-leary",
    "title": "Agentic Engineering: Working With AI, Not Just Using It — Brendan O'Leary",
    "url": "https://ai.engineer/talks/agentic-engineering-working-with-ai-not-just-using-it-brendan-o-leary",
    "videoId": "BEKc4P87XKo",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1623000,
    "speakers": [
      {
        "slug": "brendan-o-leary",
        "name": "Brendan O'Leary",
        "organization": "Kilo Code"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Brendan O'Leary explains agentic engineering as directing AI coding agents rather than passively accepting generated code. He recommends treating agents like capable but fallible junior developers, managing context with AGENTS.md and carefully selected MCP servers, restarting degraded sessions with summarized handoffs, separating research through Kilo Code Ask mode, and maintaining human control through frequent Git commits, review, build checks, and tests."
  },
  {
    "slug": "agentic-enterprise-what-your-ceo-must-know-about-ai",
    "title": "Agentic Enterprise: What Your CEO Must Know About AI",
    "url": "https://ai.engineer/talks/agentic-enterprise-what-your-ceo-must-know-about-ai",
    "videoId": "d7ds6m7fbqg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1684000,
    "speakers": [
      {
        "slug": "hubert-misztela",
        "name": "Hubert Misztela",
        "organization": "Novartis"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Novartis AI researcher Hubert Misztela explains how executives can prepare enterprises for increasingly autonomous AI agents. He outlines agent capabilities including language models, planning, tools, persistent memory, and retrieval-augmented generation; argues that organizations must map workflows, employee personas, and operational context; and describes domain assistants, team-level agents, agent swarms, simulation opportunities, and the continuing importance of ethics."
  },
  {
    "slug": "agentic-evaluations-at-scale-for-everybody",
    "title": "Agentic Evaluations at Scale — For Everybody",
    "url": "https://ai.engineer/talks/agentic-evaluations-at-scale-for-everybody",
    "videoId": "Ubwb6NzegyA",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1203000,
    "speakers": [
      {
        "slug": "nicholas-kang",
        "name": "Nicholas Kang",
        "organization": "Google DeepMind"
      },
      {
        "slug": "michael-aaron",
        "name": "Michael Aaron",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "Google DeepMind and Kaggle presenters Nicholas Kang and Michael Aaron explain why fragmented benchmarks, inconsistent evaluation harnesses, and narrow participation undermine agent evaluation. They describe community-created benchmarks informed by domain expertise, Kaggle's zero-setup Standardized Agent Exams, and Game Arena, where models compete in strategic games and appear on dynamic leaderboards."
  },
  {
    "slug": "agentic-excellence-mastering-evaluation-of-ai-agents-with-azure-ai-evaluation-sdk",
    "title": "Agentic Excellence: Mastering Evaluation of AI Agents with Azure AI Evaluation SDK",
    "url": "https://ai.engineer/talks/agentic-excellence-mastering-evaluation-of-ai-agents-with-azure-ai-evaluation-sdk",
    "videoId": "J4vPq2i0QzE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1201000,
    "speakers": [
      {
        "slug": "cedric-vidal",
        "name": "Cedric Vidal",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Microsoft Principal AI Advocate Cedric Vidal presents a layered approach to AI-agent evaluation, emphasizing that assessment should begin early in development. He contrasts red teaming with ordinary dataset-based evaluation, discusses Azure safety protections and manual model spot-checking, and demonstrates AI Toolkit in VS Code for comparing models and evaluating complete agent systems. A practical example combines Playwright MCP and GPT-4.1 to extract structured event information from webpages."
  },
  {
    "slug": "agentic-graphrag-ai-s-logical-edge",
    "title": "Agentic GraphRAG: AI’s Logical Edge",
    "url": "https://ai.engineer/talks/agentic-graphrag-ai-s-logical-edge",
    "videoId": "AvVoJBxgSQk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 927000,
    "speakers": [
      {
        "slug": "stephen-chin",
        "name": "Stephen Chin",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Neo4j developer-relations leader Stephen Chin explains how agentic GraphRAG combines graph databases, retrieval pipelines, agent orchestration, and graph-based memory to improve reasoning and reduce hallucinations and bias. He describes Neo4j’s MCP and Cypher tooling, discusses LangGraph and complementary vector retrieval, presents an enterprise knowledge-retrieval adoption example, and takes audience questions about LangChain and LangGraph."
  },
  {
    "slug": "agentic-graphrag-simplifying-retrieval-across-structured-unstructured-data-zach-blumenfeld",
    "title": "Agentic GraphRAG: Simplifying Retrieval Across Structured & Unstructured Data — Zach Blumenfeld",
    "url": "https://ai.engineer/talks/agentic-graphrag-simplifying-retrieval-across-structured-unstructured-data-zach-blumenfeld",
    "videoId": "CzM3cW6FdBs",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 925000,
    "speakers": [
      {
        "slug": "zach-blumenfeld",
        "name": "Zach Blumenfeld",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Neo4j product specialist Zach Blumenfeld demonstrates how agentic GraphRAG unifies structured and unstructured information by extracting resume data into a knowledge graph. He contrasts misleading top-k document retrieval with precise graph queries, then uses Google's Agent Development Kit, LangChain, and an MCP-connected Neo4j workflow to generate Cypher, count Python developers, compare skills, and discover collaborators as the graph schema evolves."
  },
  {
    "slug": "agentic-sdlc-at-uber-building-blocks-for-uber-s-software-factory",
    "title": "Agentic SDLC at Uber - Building Blocks for Uber’s Software Factory",
    "url": "https://ai.engineer/talks/agentic-sdlc-at-uber-building-blocks-for-uber-s-software-factory",
    "videoId": "17-YSUHo6Lk",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1106000,
    "speakers": [
      {
        "slug": "uday-kiran-medisetty",
        "name": "Uday Kiran Medisetty",
        "organization": "Uber"
      },
      {
        "slug": "adam-huda",
        "name": "Adam Huda",
        "organization": "Uber"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Uday Kiran Medisetty and Adam Huda describe Uber’s agentic software-development infrastructure, reporting that agents contribute to more than 70% of pull requests and that code output per engineer doubled year over year. They outline a centralized model gateway with SPIRE-based identity, PII redaction, and sub-100-millisecond safety controls; an MCP gateway and CLI-oriented tool access that reduced fleetwide token consumption by more than 40%; remote development environments; and context-graph-powered workflows. An end-to-end feature demonstration shows how Cortana moves from Slack-based exploration toward implementation while validating changes before they consume CI capacity."
  },
  {
    "slug": "agentic-search-for-context-engineering",
    "title": "Agentic Search for Context Engineering",
    "url": "https://ai.engineer/talks/agentic-search-for-context-engineering",
    "videoId": "ynJyIKwjonM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 3793000,
    "speakers": [
      {
        "slug": "leonie-monigatti",
        "name": "Leonie Monigatti",
        "organization": "Elastic"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Elastic developer advocate Leonie Monigatti presents a hands-on workshop on agentic search for context engineering. She explains why fixed retrieval pipelines and narrowly scoped semantic search tools fail, compares retrieval from Elasticsearch databases, local files, agent skills, memory, and the web, and examines tool descriptions, shell interfaces, context-window loading, and practical trade-offs between database queries and file search. Audience questions explore missing database information and hybrid search behavior."
  },
  {
    "slug": "agentic-security-permissions-provenance-and-the-agent-supply-chain",
    "title": "Agentic Security: Permissions, Provenance, and the Agent Supply Chain",
    "url": "https://ai.engineer/talks/agentic-security-permissions-provenance-and-the-agent-supply-chain",
    "videoId": "yWS0udrIOc8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1352000,
    "speakers": [
      {
        "slug": "steve-yegge",
        "name": "Steve Yegge",
        "organization": "Gas Town"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Steve Yegge argues that faster AI-assisted coding expands security exposure, particularly when models introduce familiar vulnerabilities or hallucinate package dependencies that attackers can replace with backdoored implementations. Drawing on Google's developer-facing testing practices and a personal project with 241 reported vulnerabilities, he emphasizes surfacing security findings directly to coding agents. Audience questions extend the discussion to agent credentials, autonomous actions, and prompt injection."
  },
  {
    "slug": "agents-are-built-at-the-fringe-getting-from-90-to-100",
    "title": "Agents are built at the fringe: getting from 90 to 100",
    "url": "https://ai.engineer/talks/agents-are-built-at-the-fringe-getting-from-90-to-100",
    "videoId": "bVNNvWq6dKo",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1250000,
    "speakers": [
      {
        "slug": "kevin-hou",
        "name": "Kevin Hou",
        "organization": "Windsurf"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Kevin Hou explains how Codeium evolved from autocomplete extensions into Windsurf, an agentic IDE built around shared user-agent context. He describes predictive workflows, continuing unfinished coding tasks, safety-aware terminal execution, operational scaling challenges, and automatic codebase, documentation, and web-context inference, illustrated with a Supabase and Next.js example."
  },
  {
    "slug": "agents-are-robots-too-what-self-driving-taught-me-about-building-agents-jesse-hu-abundant",
    "title": "Agents are Robots Too: What Self-Driving Taught Me About Building Agents — Jesse Hu, Abundant",
    "url": "https://ai.engineer/talks/agents-are-robots-too-what-self-driving-taught-me-about-building-agents-jesse-hu-abundant",
    "videoId": "qqXdLf3wy1E",
    "event": "AI Engineer Code 2025",
    "durationMs": 1057000,
    "speakers": [
      {
        "slug": "jesse-hu",
        "name": "Jesse Hu",
        "organization": "Abundant"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "Jesse Hu of Abundant applies lessons from self-driving robotics to coding agents, arguing that reliable production systems depend on embodiment, persistent state, simulation, evaluation, and offline infrastructure rather than model quality alone. He examines action-space choices through Terminal-Bench's tmux-based Terminus agent, then connects agent failures and recovery to closed-loop control, DAgger, imitation-learning distribution shift, reinforcement learning, Markov decision processes, and iterative improvement."
  },
  {
    "slug": "agents-at-scale-inside-minimax-s-model-and-the-infrastructure-behind-it",
    "title": "Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It",
    "url": "https://ai.engineer/talks/agents-at-scale-inside-minimax-s-model-and-the-infrastructure-behind-it",
    "videoId": "AVMr9PMINyo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1214000,
    "speakers": [
      {
        "slug": "dan-fu",
        "name": "Dan Fu",
        "organization": "Together AI"
      },
      {
        "slug": "olive-song",
        "name": "Olive Song",
        "organization": "MiniMax"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "In a moderated panel, Together AI's Dan Fu and MiniMax's Olive Song discuss MiniMax-M3, the partnership required to serve an open-weight multimodal model at scale, and the GPU-inference infrastructure supporting agent workloads. They examine benchmark selection, including KernelBench and OSWorld, alongside sparse attention, long-horizon agent tasks, and KV-cache challenges."
  },
  {
    "slug": "agents-building-agents",
    "title": "Agents Building Agents",
    "url": "https://ai.engineer/talks/agents-building-agents",
    "videoId": "aHhB3sjGjkI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1814000,
    "speakers": [
      {
        "slug": "alfonso-graziano",
        "name": "Alfonso Graziano",
        "organization": "Nearform"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Nearform's Alfonso Graziano explains how a coding agent can iteratively build and improve a production AI agent through spec-driven engineering, golden datasets, evaluation scorers, and human oversight. Using Mastra and an autoresearch-inspired AutoAgent loop, he demonstrates testing hypotheses, modifying prompts, tools, or code, recording results and persistent memory, and incorporating live-user feedback and observability signals such as tool usage, latency, and token consumption."
  },
  {
    "slug": "agents-don-t-do-standups-building-the-post-engineer-engineering-org",
    "title": "Agents Don't Do Standups: Building the Post-Engineer Engineering Org",
    "url": "https://ai.engineer/talks/agents-don-t-do-standups-building-the-post-engineer-engineering-org",
    "videoId": "VMemhtlsoNk",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1070000,
    "speakers": [
      {
        "slug": "mike-spitz",
        "name": "Mike Spitz",
        "organization": "PFF"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "PFF CTO Mike Spitz presents a three-month engineering case study centered on making AI agents faster rather than maximizing individual engineer output. He describes an agent-assisted workflow that turns specifications into lightweight design documents, automatically generated tickets, and pull requests; replaces standups and sprint refinement with automated coordination; standardizes recurring work through composable skills; and deploys QA agents to staging to validate acceptance criteria. Human judgment remains concentrated on specifications, design, customer outcomes, and preventing overengineering, while agent-driven remediation points toward self-healing software workflows."
  },
  {
    "slug": "agents-for-everything-else-swyx",
    "title": "Agents for Everything Else — swyx",
    "url": "https://ai.engineer/talks/agents-for-everything-else-swyx",
    "videoId": "zepu8Kk6FBQ",
    "event": "AI Engineer Europe 2026",
    "durationMs": 850000,
    "speakers": [
      {
        "slug": "swyx",
        "name": "Shawn \"swyx\" Wang",
        "organization": "AI Engineer"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "AI Engineer co-founder Shawn \"swyx\" Wang explains how a small conference team uses Cognition's Devin and Claude Cowork beyond coding: translating Figma designs into working software, eliminating setup drudgery, managing content as code, synchronizing external data, researching purchases, and automating routine knowledge work. He argues that agent-accessible APIs, CLIs, and MCP interfaces increasingly matter more than human-oriented dashboards."
  },
  {
    "slug": "agents-in-production-how-opengov-built-and-scaled-og-assist",
    "title": "Agents in Production: How OpenGov Built and Scaled OG Assist",
    "url": "https://ai.engineer/talks/agents-in-production-how-opengov-built-and-scaled-og-assist",
    "videoId": "4uFVSLgD2Q4",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1110000,
    "speakers": [
      {
        "slug": "gabe-de-mesa",
        "name": "Gabe De Mesa",
        "organization": "OpenGov"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "OpenGov engineer Gabe De Mesa explains how the company built and scaled OG Assist, an AI-agent experience embedded in government-software products. He discusses production agent architecture and tool calling, the team's use of Effect and TypeScript, A2A protocol extensions, human approval gates, long-context memory, observability, and AI-assisted development with Claude and Cursor."
  },
  {
    "slug": "agents-need-feature-flags",
    "title": "Agents Need Feature Flags",
    "url": "https://ai.engineer/talks/agents-need-feature-flags",
    "videoId": "zU4EagB311U",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1157000,
    "speakers": [
      {
        "slug": "sachin-gupta",
        "name": "Sachin Gupta",
        "organization": "eBay"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Sachin Gupta argues that AI agents need the production safeguards long used in conventional software, including segmented rollouts, canaries, rapid rollback, and kill switches. He describes six independently controlled behavior surfaces—prompts, tools, models, memory, autonomy, and sub-agents—and explains cohort-specific prompt variants, restricted tool access, privacy and compliance considerations, and per-turn runtime controls. The talk closes with operational metrics covering kill-switch activations, mitigation time, canary deltas, and audit completeness."
  },
  {
    "slug": "agents-need-more-than-a-chat",
    "title": "Agents need more than a chat",
    "url": "https://ai.engineer/talks/agents-need-more-than-a-chat",
    "videoId": "XNtkiQJ49Ps",
    "event": "AI Engineer Europe 2026",
    "durationMs": 861000,
    "speakers": [
      {
        "slug": "jacob-lauritzen",
        "name": "Jacob Lauritzen",
        "organization": "Legora"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Legora CTO Jacob Lauritzen argues that complex vertical-AI agents, particularly in legal work, require collaboration interfaces richer than linear chat. Using contract review and litigation strategy, he connects automation limits to task verifiability, then discusses guardrails, Claude Code permission tradeoffs, DAG-shaped task decomposition, reusable skills, progressive discovery, and more controllable human-agent workflows."
  },
  {
    "slug": "agents-need-receipts-not-more-tool-calls",
    "title": "Agents Need Receipts, Not More Tool Calls",
    "url": "https://ai.engineer/talks/agents-need-receipts-not-more-tool-calls",
    "videoId": "Fu45geO3zX8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1176000,
    "speakers": [
      {
        "slug": "armanas-povilionis",
        "name": "Armanas Povilionis",
        "organization": "Alithea Bio"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Armanas Povilionis of Alithea Bio argues that scientific and commercial AI-agent workflows require coordination across organizational boundaries, budgets, and verifiable records rather than merely additional local tools. He introduces Froglet, an open agent-to-agent protocol for service discovery, signed deals, execution, and receipts, and demonstrates node interaction, Docker deployment, provider-consumer roles, and integrations with MCP, Claude, OpenClaw, and NemoClaw. Potential applications include paid database access, GPU compute, and auditable scientific automation."
  },
  {
    "slug": "agents-need-receipts-not-more-tool-calls-505635",
    "title": "Agents Need Receipts, Not More Tool Calls",
    "url": "https://ai.engineer/talks/agents-need-receipts-not-more-tool-calls-505635",
    "videoId": "Q9ycQHbDdJs",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 623000,
    "speakers": [
      {
        "slug": "armanas-povilionis",
        "name": "Armanas Povilionis",
        "organization": "Alithea Bio"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Armanas Povilionis argues that automating collaborative scientific research requires verifiable cross-organizational agent interactions, not merely additional tools. He introduces Alithea Bio’s Froglet protocol for discovering services, coordinating transactions across payment and execution systems, and producing cryptographically signed chains of descriptors, offers, quotes, deals, invoices, and receipts."
  },
  {
    "slug": "agents-on-the-canvas-in-tldraw",
    "title": "Agents on the Canvas in tldraw",
    "url": "https://ai.engineer/talks/agents-on-the-canvas-in-tldraw",
    "videoId": "sPUjIBH5Cwg",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1194000,
    "speakers": [
      {
        "slug": "steve-ruiz",
        "name": "Steve Ruiz",
        "organization": "tldraw"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Steve Ruiz introduces tldraw’s React-based infinite-canvas SDK and demonstrates its progression from Make Real, which converts sketches and annotations into working interfaces, to visible AI collaborators embedded directly in the canvas. He shows how multiple Fairy agents can work concurrently across shared spatial context, arguing that on-canvas collaboration is more natural than delegating through an AI sidebar."
  },
  {
    "slug": "agents-redacted-location-address-access-redacted-location-address-and-the-future-of-machine-iden",
    "title": "Agents Access and the Future of Machine Identity",
    "url": "https://ai.engineer/talks/agents-redacted-location-address-access-redacted-location-address-and-the-future-of-machine-iden",
    "videoId": "px2e2OOS2Sk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 857000,
    "speakers": [
      {
        "slug": "nick-nisi",
        "name": "Nick Nisi",
        "organization": "WorkOS"
      },
      {
        "slug": "lizzie-siegle",
        "name": "Lizzie Siegle",
        "organization": "Cloudflare"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Nick Nisi of WorkOS and Lizzie Siegle of Cloudflare explain why AI agents need delegated identity and authorization instead of relying on developer-oriented personal access tokens. They outline Cloudflare Workers[REDACTED:location_address] bindings[REDACTED:location_address] Vectorize[REDACTED:location_address] and Durable Objects as infrastructure for stateful agents[REDACTED:location_address] then demonstrate a WorkOS-integrated MCP server connected to Claude that acts on an authenticated user's behalf and accesses an inventory tool."
  },
  {
    "slug": "agents-reported-thousands-of-bugs-how-many-were-real-ian-butler-and-nick-gregory",
    "title": "Agents reported thousands of bugs, how many were real? - Ian Butler and Nick Gregory",
    "url": "https://ai.engineer/talks/agents-reported-thousands-of-bugs-how-many-were-real-ian-butler-and-nick-gregory",
    "videoId": "wAQK7O3WGEE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1119000,
    "speakers": [
      {
        "slug": "ian-butler",
        "name": "Ian Butler",
        "organization": "Bismuth"
      },
      {
        "slug": "nick-gregory",
        "name": "Nick Gregory",
        "organization": "Bismuth"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Bismuth cofounders Ian Butler and Nick Gregory introduce SM-100, a software-maintenance benchmark built from 100 validated bugs across more than 84 public repositories. They explain why coding and security benchmarks inadequately capture real-world bug discovery, compare Bismuth and other agents across bug-finding and pull-request-review tasks, report that a basic agent loop produced a 97% false-positive rate, and argue that reliable maintenance requires stronger targeted search, program comprehension, and cross-file reasoning."
  },
  {
    "slug": "agents-vs-workflows-why-not-both",
    "title": "Agents vs Workflows: Why Not Both?",
    "url": "https://ai.engineer/talks/agents-vs-workflows-why-not-both",
    "videoId": "8SUJEqQNClw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 937000,
    "speakers": [
      {
        "slug": "sam-bhagwat",
        "name": "Sam Bhagwat",
        "organization": "Mastra"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Mastra founder Sam Bhagwat argues that autonomous agents and structured workflows are complementary architectural primitives. He examines the debate surrounding Anthropic and OpenAI guidance, criticizes cumbersome graph-oriented developer APIs using lessons from Gatsby and GraphQL, distinguishes iterative tool-using agents from dependency-ordered pipelines, and describes hybrid composition patterns including supervisor agents coordinating specialized agents."
  },
  {
    "slug": "agi-the-path-forward",
    "title": "AGI: The Path Forward",
    "url": "https://ai.engineer/talks/agi-the-path-forward",
    "videoId": "OGCG_QkCcZo",
    "event": "AI Engineer Code 2025",
    "durationMs": 956000,
    "speakers": [
      {
        "slug": "eiso-kant",
        "name": "Eiso Kant",
        "organization": "Poolside"
      },
      {
        "slug": "jason-warner",
        "name": "Jason Warner",
        "organization": "Poolside"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Poolside cofounders Jason Warner and Eiso Kant present their approach to AGI through proprietary software-engineering models trained with reinforcement learning. They demonstrate an agent in Visual Studio Code working with Ada and Rust, discuss independent model serving and Amazon Bedrock distribution, describe plans for large-scale GB300 compute and vertically integrated data centers, and recount their shared history around GitHub and early AI-assisted coding."
  },
  {
    "slug": "ai-agents-for-performance-ship-faster-pay-less",
    "title": "AI Agents for Performance: Ship Faster, Pay Less",
    "url": "https://ai.engineer/talks/ai-agents-for-performance-ship-faster-pay-less",
    "videoId": "CgsWxRUY5Eo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2019000,
    "speakers": [
      {
        "slug": "rajat-shah",
        "name": "Rajat Shah",
        "organization": "Netflix"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Netflix staff software engineer Rajat Shah explains how AI agents can analyze CPU profiles and call stacks, inspect the exact Git revision deployed in production, and identify expensive code paths. He describes a quadratic-time pattern responsible for 8.8% of CPU during profiling, measured production savings after optimization, and a shared anti-pattern catalog that gives coding and review agents reusable knowledge across services and earlier stages of software development."
  },
  {
    "slug": "ai-agents-meet-test-driven-development",
    "title": "AI Agents, Meet Test Driven Development",
    "url": "https://ai.engineer/talks/ai-agents-meet-test-driven-development",
    "videoId": "U3MVU6JpocU",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1750000,
    "speakers": [
      {
        "slug": "anita-kirkovska",
        "name": "Anita Kirkovska",
        "organization": "Vellum"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Vellum's Anita Kirkovska explains how test-driven development can make nondeterministic AI agents and agentic workflows reliable in production. She surveys reasoning models, chain-of-thought prompting, RAG, GraphRAG, agentic RAG, and ReAct; emphasizes domain-expert participation, observability, fallback models, staging, and regression checks; discusses levels of agent behavior and feedback-memory loops; and introduces Vellum's code-oriented Workflows SDK."
  },
  {
    "slug": "ai-and-human-whiteboarding-partnership",
    "title": "AI and Human Whiteboarding Partnership",
    "url": "https://ai.engineer/talks/ai-and-human-whiteboarding-partnership",
    "videoId": "aopgVJBQC0o",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1019000,
    "speakers": [
      {
        "slug": "christopher-chedeau",
        "name": "Christopher Chedeau"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Excalidraw creator Christopher Chedeau recounts prototyping a hand-drawn virtual whiteboard with Zwibbler as inspiration and Rough.js, then explains how pandemic-era remote collaboration exposed opportunities to surpass physical whiteboard constraints. He contrasts superficial AI additions with human-centered whiteboarding workflows, discusses an Anthropic Claude coding-agent demonstration, and shows how AI-assisted diagrams remain directly editable within Excalidraw."
  },
  {
    "slug": "ai-changes-nothing-dax-raad-opencode",
    "title": "AI changes *Nothing* — Dax Raad, OpenCode",
    "url": "https://ai.engineer/talks/ai-changes-nothing-dax-raad-opencode",
    "videoId": "o3gmwzo-Mik",
    "event": "AI Engineer Code 2025",
    "durationMs": 999000,
    "speakers": [
      {
        "slug": "dax-raad",
        "name": "Dax Raad",
        "organization": "OpenCode"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "OpenCode contributor Dax Raad argues that AI does not change the fundamentals of building a successful product. Using a three-stage product funnel, he examines distinctive creative marketing, frictionless onboarding that delivers a compelling aha moment, and durable product depth that serves both beginners and demanding users. His central argument is that taste, creativity, deep understanding, and difficult execution remain essential even when AI accelerates coding."
  },
  {
    "slug": "ai-consulting-in-practice-nathaniel-whittemore-nlw-superintelligent",
    "title": "AI Consulting in Practice — Nathaniel Whittemore (NLW), Superintelligent",
    "url": "https://ai.engineer/talks/ai-consulting-in-practice-nathaniel-whittemore-nlw-superintelligent",
    "videoId": "ehQFj6VmuI8",
    "event": "AI Engineer Code 2025",
    "durationMs": 1098000,
    "speakers": [
      {
        "slug": "nathaniel-whittemore-nlw",
        "name": "Nathaniel Whittemore (NLW)",
        "organization": "Superintelligent"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Nathaniel Whittemore (NLW), host of The AI Daily Brief and CEO of Superintelligent, examines where enterprises are finding practical value from AI. He contrasts rising AI-agent deployment with persistent difficulty scaling beyond pilots, discusses executive spending and ROI expectations, and presents an audience-sourced benchmarking study covering roughly 1,000 organizations and 3,500 reported use cases. Findings highlight measurable time savings, differences between adoption leaders and laggards, organization-size effects, and the selection bias inherent in a self-reporting AI-focused audience."
  },
  {
    "slug": "ai-copilots-for-tech-architecture-the-highest-roi-use-case-you-re-not-building-boris-bogatin-and",
    "title": "AI Copilots for Tech Architecture: The Highest-ROI Use Case You’re Not Building — Boris Bogatin and Toufic Boubez, Catio",
    "url": "https://ai.engineer/talks/ai-copilots-for-tech-architecture-the-highest-roi-use-case-you-re-not-building-boris-bogatin-and",
    "videoId": "QRWdapxMdSY",
    "event": "AI Engineer Code 2025",
    "durationMs": 1649000,
    "speakers": [
      {
        "slug": "boris-bogatin",
        "name": "Boris Bogatin",
        "organization": "Catio"
      },
      {
        "slug": "toufic-boubez",
        "name": "Toufic Boubez",
        "organization": "Catio"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Catio co-founders Boris Bogatin and Toufic Boubez argue that AI copilots should improve architecture decisions, not just coding productivity. They identify three enterprise challenges: visibility across technology estates, ROI-backed prioritization, and scalable guidance for developers. Their approach combines live architecture mapping through Catio Stacks, business-aware recommendations, collaborative multi-agent problem-solving, and measurable feedback loops to align technical decisions with organizational objectives."
  },
  {
    "slug": "ai-didn-t-kill-the-web-it-moved-in-olivier-leplus-aws-yohan-lasorsa-microsoft",
    "title": "AI Didn’t Kill the Web, It Moved in! — Olivier Leplus (AWS) & Yohan Lasorsa (Microsoft)",
    "url": "https://ai.engineer/talks/ai-didn-t-kill-the-web-it-moved-in-olivier-leplus-aws-yohan-lasorsa-microsoft",
    "videoId": "XZ0boOjtbNo",
    "event": "AI Engineer Europe 2026",
    "durationMs": 3175000,
    "speakers": [
      {
        "slug": "olivier-leplus",
        "name": "Olivier Leplus",
        "organization": "AWS"
      },
      {
        "slug": "yohan-lasorsa",
        "name": "Yohan Lasorsa",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Microsoft developer advocate Yohan Lasorsa and AWS developer advocate Olivier Leplus demonstrate how AI is reshaping the web-development lifecycle. Using an example e-commerce application, they show coding agents implementing GitHub issues with reusable SKILL.md instructions and GitHub CLI, discuss Playwright and browser-assisted debugging, demonstrate browser-native AI capabilities including summarization and text correction, and explain how agent-readable documentation and MCP/WebMCP tools make websites usable by both people and AI agents."
  },
  {
    "slug": "ai-driven-multi-document-correlation-for-enterprise-financial-compliance-and-fraud-detection",
    "title": "AI-Driven Multi-Document Correlation for Enterprise Financial Compliance and Fraud Detection",
    "url": "https://ai.engineer/talks/ai-driven-multi-document-correlation-for-enterprise-financial-compliance-and-fraud-detection",
    "videoId": "Iwe_RY-fYgI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1140000,
    "speakers": [
      {
        "slug": "varsha-shah",
        "name": "Varsha Shah",
        "organization": "Independent Researcher"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Varsha Shah presents a research-driven framework for detecting financial compliance risks and fraud across payroll, tax, procurement, and transaction records. The approach combines graph-based entity correlation, probabilistic risk modeling, and cross-jurisdictional normalization, then incorporates investigator feedback to refine risk scoring. Shah reports evaluation on approximately three million financial records spanning five years and four jurisdictions, with an F1 score of 0.89, and outlines integration with enterprise and audit systems."
  },
  {
    "slug": "ai-engineering-101",
    "title": "AI Engineering 101",
    "url": "https://ai.engineer/talks/ai-engineering-101",
    "videoId": "C0ZUdFg-iTo",
    "event": "AI Engineer Summit 2023",
    "durationMs": 10944000,
    "speakers": [
      {
        "slug": "noah-hein",
        "name": "Noah Hein",
        "organization": "Latent Space University"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Noah Hein leads an introductory, hands-on AI engineering workshop based on Latent Space University material. Participants configure a Telegram bot and development environment, connect to the OpenAI API, and discuss conversation context windows, tokenization, model temperature, iterative prompting, AI-assisted code generation, and image-generation applications including RoomGPT."
  },
  {
    "slug": "ai-engineering-201-inference",
    "title": "AI Engineering 201: Inference",
    "url": "https://ai.engineer/talks/ai-engineering-201-inference",
    "videoId": "N7lJY5IKVLE",
    "event": "AI Engineer Summit 2023",
    "durationMs": 6196000,
    "speakers": [
      {
        "slug": "charles-frye",
        "name": "Charles Frye"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Charles Frye presents the inference-focused first portion of AI Engineering 201, examining why model execution requires costly accelerators, how proprietary APIs compare with open models such as Llama and Mistral 7B, and the tradeoffs among on-device, hosted, serverless, and self-managed deployment. Audience questions address language support, model memory, browser and mobile constraints, and alternative accelerators; the workshop also covers latency, request batching, throughput, Kubernetes, and inference-serving platforms."
  },
  {
    "slug": "ai-engineering-201-the-rest-of-the-owl",
    "title": "AI Engineering 201: The Rest of the Owl",
    "url": "https://ai.engineer/talks/ai-engineering-201-the-rest-of-the-owl",
    "videoId": "zl4EdALzktU",
    "event": "AI Engineer Summit 2023",
    "durationMs": 3417000,
    "speakers": [
      {
        "slug": "charles-frye",
        "name": "Charles Frye"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Charles Frye’s interactive AI Engineering 201 workshop examines the engineering required around model inference to build useful AI products. It covers language user interfaces and their historical predecessors, retrieval and contextual grounding, typed and structured model outputs, function calling and tool-use patterns including TypeChat and Instructor, and production testing, monitoring, observability, and evaluation."
  },
  {
    "slug": "ai-engineering-with-the-google-gemini-2-5-model-family",
    "title": "AI Engineering with the Google Gemini 2.5 Model Family",
    "url": "https://ai.engineer/talks/ai-engineering-with-the-google-gemini-2-5-model-family",
    "videoId": "zK9lYrLbjSg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 6291000,
    "speakers": [
      {
        "slug": "philipp-schmid",
        "name": "Philipp Schmid",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Google DeepMind developer-relations engineer Philipp Schmid leads a hands-on Gemini 2.5 workshop covering Flash and Pro models, API authentication, the accompanying GitHub notebooks, multimodal inputs, conversational state management, and agent tooling. Demonstrations and audience questions explore local and remote MCP servers, DeepWiki repository access, document processing, video understanding, prompt caching, and source citations."
  },
  {
    "slug": "ai-engineering-without-borders",
    "title": "AI Engineering Without Borders",
    "url": "https://ai.engineer/talks/ai-engineering-without-borders",
    "videoId": "eDr0m6R7fI4",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 632000,
    "speakers": [
      {
        "slug": "swyx",
        "name": "swyx",
        "organization": "Latent.Space"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "swyx argues that AI engineering should not be constrained by rigid professional definitions or artificial boundaries between technical disciplines. He describes multilingual and multimodal model capabilities, distinguishes emerging AI-engineer archetypes, and connects RAG, code generation, agents, open models, evaluation, team strategy, and community. The talk also considers software-engineering definitions, human interaction constraints, and improving context."
  },
  {
    "slug": "ai-frontiers-in-trust-and-safety-combatting-multifaceted-harm-on-tinder-at-scale",
    "title": "AI Frontiers in Trust and Safety: Combatting Multifaceted Harm on Tinder at Scale",
    "url": "https://ai.engineer/talks/ai-frontiers-in-trust-and-safety-combatting-multifaceted-harm-on-tinder-at-scale",
    "videoId": "kwnCvA9l-TY",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 876000,
    "speakers": [
      {
        "slug": "vibhor-kumar",
        "name": "Vibhor Kumar",
        "organization": "Tinder"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Tinder engineer Vibhor Kumar explains how generative AI amplifies spam, scams, and other trust-and-safety risks while also enabling stronger harm detection. He describes fine-tuning open language models, generating classification datasets through hybrid human/LLM workflows, adapting to adversarial fraud, and serving multiple fine-tuned adapters efficiently with LoRAX and automated retraining."
  },
  {
    "slug": "ai-is-the-world-s-largest-relationship-therapist-clay-cockrell-tony-fabrikant-couplework-ai",
    "title": "AI is the World’s largest Relationship Therapist — Clay Cockrell & Tony Fabrikant, CoupleWork AI",
    "url": "https://ai.engineer/talks/ai-is-the-world-s-largest-relationship-therapist-clay-cockrell-tony-fabrikant-couplework-ai",
    "videoId": "yoONZwV2smc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1003000,
    "speakers": [
      {
        "slug": "clay-cockrell",
        "name": "Clay Cockrell",
        "organization": "CoupleWork AI"
      },
      {
        "slug": "tony-fabrikant",
        "name": "Tony Fabrikant",
        "organization": "CoupleWork AI"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "CoupleWork co-founders Clay Cockrell and Tony Fabrikant argue that relationship-focused AI must prioritize clinical outcomes over engagement and agreement. Cockrell identifies sycophancy, inadequate risk escalation, and weak privacy protections as hazards, describes foundations in Gottman research and Emotionally Focused Therapy, and presents Maxine as a clinician-informed relationship coach designed to recognize dangerous situations. Fabrikant closes by urging engineers to partner with clinicians and meet meaningful clinical standards."
  },
  {
    "slug": "ai-kernel-generation-what-s-working-what-s-not-what-s-next",
    "title": "AI Kernel Generation: What's Working, What's Not, What's Next",
    "url": "https://ai.engineer/talks/ai-kernel-generation-what-s-working-what-s-not-what-s-next",
    "videoId": "6guQG_tGt0o",
    "event": "AI Engineer Code 2025",
    "durationMs": 1155000,
    "speakers": [
      {
        "slug": "natalie-serrino",
        "name": "Natalie Serrino",
        "organization": "Gimlet Labs"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Gimlet Labs co-founder Natalie Serrino explains how AI agents generate and optimize GPU kernels to run PyTorch workloads efficiently across heterogeneous hardware. She demonstrates an H100-targeted workflow and presents Apple M4/Metal results on KernelBench v0.1, reporting approximately 24–25% average speedup, a 40% kernel-fusion example, and an 80% operator-optimization example. She emphasizes profiling, correctness checks, hardware-specific context, and continued human supervision."
  },
  {
    "slug": "ai-music-generation-from-prompt-to-production",
    "title": "AI Music Generation: From Prompt to Production",
    "url": "https://ai.engineer/talks/ai-music-generation-from-prompt-to-production",
    "videoId": "CoaL4JZKsWI",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 3274000,
    "speakers": [
      {
        "slug": "phlo-young",
        "name": "Phlo Young"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Phlo Young leads an interactive beginner-oriented workshop on creating and refining AI-generated music. He introduces Boomy, Suno, and Udio; demonstrates AI vocal examples and ElevenLabs text-to-music; gathers audience preferences for prompting techniques; and discusses music distribution, extracting stems, mixing, mastering, and using ChatGPT to develop music-generation prompts."
  },
  {
    "slug": "ai-on-your-lakehouse-context-comes-in-shapes-not-queries",
    "title": "AI on Your Lakehouse: Context Comes in Shapes, Not Queries",
    "url": "https://ai.engineer/talks/ai-on-your-lakehouse-context-comes-in-shapes-not-queries",
    "videoId": "kRkcNOsRyYg",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 7150000,
    "speakers": [
      {
        "slug": "zach-blumenfeld",
        "name": "Zach Blumenfeld",
        "organization": "Neo4j"
      },
      {
        "slug": "ben-squire",
        "name": "Ben Squire",
        "organization": "Neo4j"
      },
      {
        "slug": "ryan-knight",
        "name": "Ryan Knight",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Neo4j AI Research Engineer Zach Blumenfeld leads a hands-on workshop showing how graph-shaped context makes lakehouse-connected AI agents more reliable. Using GraphAcademy, GitHub Codespaces, Claude Code, BigQuery, Neo4j CLI, and agent-generated Cypher, participants build reusable document-outline trees, thematic communities, and connection paths that federate documents with warehouse records. The session covers semantic expansion, Lucene-backed search, idempotent loading, and finding undocumented field problems; the patterns extend to Databricks and Snowflake."
  },
  {
    "slug": "ai-pipelines-and-agents-in-pure-typescript-with-mastra-ai",
    "title": "AI Pipelines and Agents in Pure TypeScript with Mastra.ai",
    "url": "https://ai.engineer/talks/ai-pipelines-and-agents-in-pure-typescript-with-mastra-ai",
    "videoId": "FWlRHPZWyHE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 6674000,
    "speakers": [
      {
        "slug": "nick-nisi",
        "name": "Nick Nisi",
        "organization": "WorkOS"
      },
      {
        "slug": "zack-proser",
        "name": "Zack Proser",
        "organization": "WorkOS"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "WorkOS presenters Nick Nisi and Zack Proser lead a hands-on workshop building an AI-powered meme generator with Mastra and TypeScript. They demonstrate composable agent workflows, structured outputs, OpenAI and Imgflip integrations, Git-based workshop checkpoints, and Model Context Protocol tooling, while attendees and an additional Mastra participant contribute questions and implementation discussion."
  },
  {
    "slug": "ai-platform-engineering",
    "title": "AI Platform Engineering",
    "url": "https://ai.engineer/talks/ai-platform-engineering",
    "videoId": "5qNXdLbEdew",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1698000,
    "speakers": [
      {
        "slug": "patrick-debois",
        "name": "Patrick Debois"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Patrick Debois applies lessons from the emergence of DevOps to explain why organizations need dedicated AI platform teams as generative AI moves beyond isolated data-science initiatives. He describes enabling application developers through shared infrastructure, Kubernetes and CloudOps coordination, caching, feedback systems, local experimentation, engineering rigor, risk assessment, and centralized governance with self-service access."
  },
  {
    "slug": "ai-powered-entomology-lessons-from-millions-of-ai-code-reviews",
    "title": "AI-powered entomology: Lessons from millions of AI code reviews",
    "url": "https://ai.engineer/talks/ai-powered-entomology-lessons-from-millions-of-ai-code-reviews",
    "videoId": "TswQeKftnaw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 621000,
    "speakers": [
      {
        "slug": "tomas-reimers",
        "name": "Tomas Reimers",
        "organization": "Graphite"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Graphite co-founder Tomas Reimers explains how its Diamond AI code-review product detects real pull-request bugs while avoiding hallucinated, pedantic, or otherwise unwelcome feedback. Drawing on an analysis of 10,000 review comments, he presents a two-axis framework separating what language models can reliably identify from what developers actually want to receive, highlights undocumented team knowledge as a persistent limitation, and describes using developer votes to evaluate model and product quality."
  },
  {
    "slug": "ai-red-teaming-agent-azure-ai-foundry-nagkumar-arkalgud-keiji-kanazawa-microsoft",
    "title": "AI Red Teaming Agent: Azure AI Foundry — Nagkumar Arkalgud & Keiji Kanazawa, Microsoft",
    "url": "https://ai.engineer/talks/ai-red-teaming-agent-azure-ai-foundry-nagkumar-arkalgud-keiji-kanazawa-microsoft",
    "videoId": "JhJKgRAmfIU",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1171000,
    "speakers": [
      {
        "slug": "nagkumar-arkalgud",
        "name": "Nagkumar Arkalgud",
        "organization": "Microsoft"
      },
      {
        "slug": "keiji-kanazawa",
        "name": "Keiji Kanazawa",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Microsoft presenters Keiji Kanazawa and Nagkumar Arkalgud explain how adversarial prompts can circumvent model safeguards and why AI agents require systematic security and safety evaluation. They introduce Microsoft's AI Red Team, the open-source PyRIT framework, and its integration into Azure AI Foundry through an evaluation SDK and hosted reporting. Demonstrations include testing a locally running Ollama model and scanning an Azure OpenAI configuration, followed by an audience question about guardrails."
  },
  {
    "slug": "ai-s-jurassic-park-period",
    "title": "AI’s Jurassic Park Period",
    "url": "https://ai.engineer/talks/ai-s-jurassic-park-period",
    "videoId": "1lgFGaHoGq8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1302000,
    "speakers": [
      {
        "slug": "aaron-stanley",
        "name": "Aaron Stanley",
        "organization": "dbt Labs"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "dbt Labs security leader Aaron Stanley compares contemporary AI agents to an inexperienced forensic investigator who damaged evidence while working around a constraint. He argues that agents can violate explicit instructions while appearing compliant, so sandboxing, egress filters, audit trails, and telemetry must be complemented by corrigibility, meaningful human intervention, and enforceable constraints, particularly for high-risk AI subject to EU AI Act oversight requirements. The presentation concludes with audience questions."
  },
  {
    "slug": "ai-sdk-v6",
    "title": "AI SDK v6",
    "url": "https://ai.engineer/talks/ai-sdk-v6",
    "videoId": "wflNENRSUb4",
    "event": "AI Engineer Europe 2026",
    "durationMs": 4133000,
    "speakers": [
      {
        "slug": "nico-albanese",
        "name": "Nico Albanese",
        "organization": "Vercel"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "In this hands-on AI SDK v6 workshop, Vercel's Nico Albanese configures a Next.js project with the Vercel CLI, OIDC, AI Gateway, and an OpenAI provider; then demonstrates multi-step agent tooling and Vercel Sandbox integration. He explains how giving agents an isolated filesystem and persistent working context improves follow-through, supports reusable Python-generated tools and memory, and enables durable coding-agent workflows."
  },
  {
    "slug": "ai-system-design-from-idea-to-production",
    "title": "AI System Design: From Idea to Production",
    "url": "https://ai.engineer/talks/ai-system-design-from-idea-to-production",
    "videoId": "T0HhO4YtTfE",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1733000,
    "speakers": [
      {
        "slug": "apoorva-joshi",
        "name": "Apoorva Joshi",
        "organization": "MongoDB"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "MongoDB developer advocate Apoorva Joshi presents a repeatable framework for taking AI systems from initial requirements to production, using a hypothetical health-insurance claims-review application. She covers defining the business problem and success metrics, identifying human-review and infrastructure constraints, selecting retrieval-augmented generation where appropriate, establishing guardrails and evaluation criteria, and monitoring production signals such as human override rates and review time."
  },
  {
    "slug": "ai-that-pays-lessons-from-revenue-cycle",
    "title": "AI That Pays: Lessons from Revenue Cycle",
    "url": "https://ai.engineer/talks/ai-that-pays-lessons-from-revenue-cycle",
    "videoId": "TquUsN1QsWs",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1099000,
    "speakers": [
      {
        "slug": "nathan-wan",
        "name": "Nathan Wan",
        "organization": "Ensemble Health Partners"
      }
    ],
    "topics": [
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Nathan Wan, head of AI at Ensemble Health Partners, argues that healthcare revenue-cycle management is an overlooked opportunity for applied AI because manual claims processes, payment delays, and denials increase financial pressure on hospitals. Drawing on experience in speech recognition and language modeling, he contrasts clinical AI with operational applications, explains provider-payer friction, and describes how longitudinal revenue-cycle data can support denial prediction and correction. He also outlines how generative AI can accelerate review of medical records and payer policies and help draft denial appeal letters."
  },
  {
    "slug": "ai-tools-for-forward-deployed-engineering",
    "title": "AI tools for Forward Deployed Engineering",
    "url": "https://ai.engineer/talks/ai-tools-for-forward-deployed-engineering",
    "videoId": "l0FLhNqBOic",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1223000,
    "speakers": [
      {
        "slug": "vasuman-moza",
        "name": "Vasuman Moza",
        "organization": "Varick Agents"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Varick Agents CEO Vasuman Moza argues that enterprise AI’s principal bottleneck is understanding and redesigning customer-specific workflows, not executing isolated tasks. He explains how forward-deployed engineers embed with clients, map finance and operational processes, and build agents over existing systems of record without scaling headcount proportionally. An additional, publicly unidentified platform-team presenter describes internal tooling, including an engagement agent for forward-deployed engineers, and discusses pragmatic infrastructure choices and frontier-model verbosity."
  },
  {
    "slug": "alphalab-autonomous-multi-agent-research-across-optimization-domains-with-frontier-llms-brendan",
    "title": "ALPHALAB: Autonomous Multi-Agent Research Across Optimization Domains with Frontier LLMs — Brendan Rappazzo",
    "url": "https://ai.engineer/talks/alphalab-autonomous-multi-agent-research-across-optimization-domains-with-frontier-llms-brendan",
    "videoId": "kiqubc5b5Yo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1207000,
    "speakers": [
      {
        "slug": "brendan-rappazzo",
        "name": "Brendan Rappazzo",
        "organization": "Morgan Stanley"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Brendan Rappazzo presents Morgan Stanley’s AlphaLab, a custom, provider-agnostic multi-agent harness for automating quantitative research from a dataset and natural-language objective. He explains the progression from the publicly documented AlphaLab 1.0 toward AlphaLab 2.0, emphasizing domain research, trustworthy evaluation construction, and large-scale worker-agent experimentation coordinated through Kanban-style workflows and Slurm-managed GPU clusters. The talk highlights enterprise research context, frontier coding models, evaluation failures, and iterative improvements to autonomous experimentation."
  },
  {
    "slug": "always-on-agents-run-production-without-the-on-call-tax",
    "title": "Always-on agents run production without the on-call tax",
    "url": "https://ai.engineer/talks/always-on-agents-run-production-without-the-on-call-tax",
    "videoId": "vSx5IULvBns",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1496000,
    "speakers": [
      {
        "slug": "justin-smith",
        "name": "Justin Smith",
        "organization": "Resolve AI"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Resolve AI founding product engineer Justin Smith argues that AI-assisted coding increases deployment velocity and production complexity without eliminating operational bottlenecks. He describes always-on background agents that combine production context, reasoning, and actions to monitor deployments, investigate issues such as P99 drift, and reduce recurring on-call work while running persistently in cloud sandboxes."
  },
  {
    "slug": "amazon-agi",
    "title": "Agents, codebases, and teams: what it actually takes to ship together",
    "url": "https://ai.engineer/talks/amazon-agi",
    "videoId": "aeTb5BdmTTc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1017000,
    "speakers": [
      {
        "slug": "aditya-khandelwal",
        "name": "Aditya Khandelwal",
        "organization": "Amazon AGI Lab"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Aditya Khandelwal argues that getting engineering teams to ship reliably with AI coding agents is a leadership and organizational challenge, not merely an individual tooling problem. Drawing on experience leading a ten-person team, he describes production incidents, uneven pull-request review burdens, progressive disclosure, shared coding practices, a ShipIt skill that automates pull-request preparation and CI remediation, and closed-loop workflows using agentic reviews and nightly code gardening."
  },
  {
    "slug": "amp-code-next-generation-ai-coding",
    "title": "Amp Code: Next-Generation AI Coding",
    "url": "https://ai.engineer/talks/amp-code-next-generation-ai-coding",
    "videoId": "gvIAkmZUEZY",
    "event": "AI Engineer Code 2025",
    "durationMs": 1111000,
    "speakers": [
      {
        "slug": "beyang-liu",
        "name": "Beyang Liu",
        "organization": "Amp Code / Sourcegraph"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Beyang Liu introduces Amp as an opinionated frontier coding agent spanning terminal and editor workflows. He explains its custom terminal interface, integrations with editors including Emacs, Neovim, JetBrains, and VS Code derivatives, and preference for specialized built-in tools over broad MCP integrations. The talk examines context exhaustion, Finder and other specialized subagents, modern terminal rendering, and how developers must adapt their craft alongside increasingly capable coding agents."
  },
  {
    "slug": "analyzing-10-000-sales-calls-with-ai-in-2-weeks",
    "title": "Analyzing 10,000 Sales Calls with AI in 2 Weeks",
    "url": "https://ai.engineer/talks/analyzing-10-000-sales-calls-with-ai-in-2-weeks",
    "videoId": "dvft0Gp9sEE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 589000,
    "speakers": [
      {
        "slug": "charlie-guo",
        "name": "Charlie Guo",
        "organization": "Pulley"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Charlie Guo describes analyzing 10,000 Pulley sales-call transcripts within two weeks to identify more precise customer profiles and extract actionable customer intelligence. After cheaper models produced misleading classifications, the team chose Claude 3.5 Sonnet, added structured JSON outputs and source citations, and used prompt caching and longer single-pass generation to reduce analysis costs from $5,000 to $500. The resulting system evolved into a reusable internal tool with search, filters, and exports."
  },
  {
    "slug": "anchoring-enterprise-genai-with-knowledge-graphs",
    "title": "Anchoring Enterprise GenAI with Knowledge Graphs",
    "url": "https://ai.engineer/talks/anchoring-enterprise-genai-with-knowledge-graphs",
    "videoId": "OpVkWc3YnFc",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1259000,
    "speakers": [
      {
        "slug": "stephen-chin",
        "name": "Stephen Chin",
        "organization": "Neo4j"
      },
      {
        "slug": "jonathan-lowe",
        "name": "Jonathan Lowe",
        "organization": "Pfizer"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Neo4j’s Stephen Chin and Pfizer’s Jonathan Lowe discuss grounding enterprise generative AI in knowledge graphs and GraphRAG while securing organizational support for production deployment. Drawing on life-sciences manufacturing and knowledge-transfer challenges, they cover graph-based document chunking, contextual retrieval, governance, project failure risks, architecture costs, and communicating concrete business value to executives."
  },
  {
    "slug": "announcing-the-ai-engineer-network",
    "title": "Announcing the AI Engineer Network",
    "url": "https://ai.engineer/talks/announcing-the-ai-engineer-network",
    "videoId": "dQmseZ6kz8w",
    "event": "AI Engineer Summit 2023",
    "durationMs": 464000,
    "speakers": [
      {
        "slug": "benjamin-dunphy",
        "name": "Benjamin Dunphy",
        "organization": "Software 3.0, LLC"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Benjamin Dunphy opens AI Engineer Summit 2023 by welcoming attendees and recognizing ecosystem sponsors including AutoGPT, Supabase, Fixie, Microsoft, and Cloudflare. He introduces Network, the conference mobile application, announces LLM-powered attendee matching through its Generative Matching Algorithm, points listeners to its open-source repository, and invites co-host Swyx to the stage."
  },
  {
    "slug": "anthropic-for-vps-of-ai",
    "title": "Anthropic for VPs of AI",
    "url": "https://ai.engineer/talks/anthropic-for-vps-of-ai",
    "videoId": "VhPfM_aGBVc",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1255000,
    "speakers": [
      {
        "slug": "alexander-bricken",
        "name": "Alexander Bricken",
        "organization": "Anthropic"
      },
      {
        "slug": "joe-bayley",
        "name": "Joe Bayley",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Anthropic’s Alexander Bricken and Joe Bayley outline practical enterprise AI adoption, introducing Claude 3.5 Sonnet and coding evaluations before explaining interpretability through semantic feature activations and the Golden Gate Claude steering demonstration. They discuss Model Context Protocol, Intercom’s Fin customer-service deployment, and implementation practices centered on LLMOps, use-case-specific metrics, rigorous evaluation, and solving concrete business problems."
  },
  {
    "slug": "anthropic-s-applied-ai-team-on-the-evolution-of-agentic-surfaces",
    "title": "Anthropic's Applied AI team on the Evolution of Agentic Surfaces",
    "url": "https://ai.engineer/talks/anthropic-s-applied-ai-team-on-the-evolution-of-agentic-surfaces",
    "videoId": "K0X9QDRkIdg",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1884000,
    "speakers": [
      {
        "slug": "gagan-bhat",
        "name": "Gagan Bhat",
        "organization": "Anthropic"
      },
      {
        "slug": "isabella-kai-he",
        "name": "Isabella Kai He",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Anthropic Applied AI presenters Gagan Bhat and Isabella Kai He trace agent-building surfaces from the Messages API and Claude Agent SDK to Claude Managed Agents. They explain why harness assumptions must evolve with model capabilities, describe separating agent reasoning from sandboxed tool execution, and show how durable session logs, isolated environments, observability, and credential vaults support reliable production agents."
  },
  {
    "slug": "anthropic-s-cca-exam-as-a-field-guide-for-agentic-engineering",
    "title": "Anthropic's CCA Exam as a Field-Guide for Agentic Engineering",
    "url": "https://ai.engineer/talks/anthropic-s-cca-exam-as-a-field-guide-for-agentic-engineering",
    "videoId": "Z-c11pV_uvU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1208000,
    "speakers": [
      {
        "slug": "frank-coyle",
        "name": "Frank Coyle",
        "organization": "UCAL Berkeley"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "UC Berkeley instructor Frank Coyle presents Anthropic’s Claude Certified Architect certification as a practical guide to building production-oriented AI agents. He surveys exam domains including agentic architecture, Claude Code, prompt engineering, JSON, tool design and Model Context Protocol integration, then explains orchestration loops, application-executed tools, stop reasons, token exhaustion, CLAUDE.md instructions, specialized subagents and context isolation."
  },
  {
    "slug": "any-to-any-building-native-multimodal-agents",
    "title": "Any-to-Any: Building Native Multimodal Agents",
    "url": "https://ai.engineer/talks/any-to-any-building-native-multimodal-agents",
    "videoId": "GIRpQEfYf3U",
    "event": "AI Engineer Europe 2026",
    "durationMs": 981000,
    "speakers": [
      {
        "slug": "patrick-lober",
        "name": "Patrick Löber",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Google DeepMind's Patrick Löber explains how to build any-to-any multimodal agents using Gemini as a reasoning model that ingests documents, images, audio, and video and invokes specialized image, speech, and real-time interaction capabilities. Using a NotebookLM-style application, he outlines file uploads, Gemini SDK calls, agentic tool orchestration, native image generation, configurable speech, and related capabilities including multimodal embeddings, Gemma, and Veo."
  },
  {
    "slug": "architecting-agent-memory-principles-patterns-and-best-practices",
    "title": "Architecting Agent Memory: Principles, Patterns, and Best Practices",
    "url": "https://ai.engineer/talks/architecting-agent-memory-principles-patterns-and-best-practices",
    "videoId": "W2HVdB4Jbjs",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1057000,
    "speakers": [
      {
        "slug": "richmond-alake",
        "name": "Richmond Alake",
        "organization": "MongoDB"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "MongoDB’s Richmond Alake explains why dependable AI agents require persistent memory alongside perception, LLM-based reasoning, and tool use. He traces the progression from chatbots and RAG to agentic systems, outlines conversational, entity, working, short-term, and long-term memory, and describes MongoDB-backed storage and retrieval patterns for agentic RAG. He closes by connecting agent-memory research to neuroscience, convolutional neural networks, and Voyage AI."
  },
  {
    "slug": "architecting-and-testing-controllable-agents",
    "title": "Architecting and Testing Controllable Agents",
    "url": "https://ai.engineer/talks/architecting-and-testing-controllable-agents",
    "videoId": "ib-wTAvCZqg",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 8514000,
    "speakers": [
      {
        "slug": "lance-martin",
        "name": "Lance Martin",
        "organization": "LangChain"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Lance Martin of LangChain leads a hands-on workshop on building controllable LLM agents with LangGraph and evaluating them with LangSmith. He contrasts fixed chains with model-directed control flow, demonstrates tool-calling loops and multi-agent notebooks, and examines answer validation, agentic RAG state, document chunking, retrieval precision-recall tradeoffs, long-context limitations, and production constraints."
  },
  {
    "slug": "are-mcps-overhyped-a-rant-about-mcps",
    "title": "Are MCPs Overhyped? A Rant about MCPs",
    "url": "https://ai.engineer/talks/are-mcps-overhyped-a-rant-about-mcps",
    "videoId": "tOou_GJ9Ddk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 449000,
    "speakers": [
      {
        "slug": "henry-mao",
        "name": "Henry Mao",
        "organization": "Smithery"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Smithery founder Henry Mao argues that stronger models and the Model Context Protocol alone do not make AI agents practically useful. He outlines MCP ecosystem problems including fragmented discovery, installation friction, insecure servers, agentic payments, hosting, developer tooling, distribution, and observability, then presents Smithery as an orchestration gateway and demonstrates an agent connecting GitHub work to Linear."
  },
  {
    "slug": "arrakis-how-to-build-an-ai-sandbox-from-scratch",
    "title": "Arrakis: How To Build An AI Sandbox From Scratch",
    "url": "https://ai.engineer/talks/arrakis-how-to-build-an-ai-sandbox-from-scratch",
    "videoId": "wsFd22SL1s8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2418000,
    "speakers": [
      {
        "slug": "abhishek-bhardwaj",
        "name": "Abhishek Bhardwaj",
        "organization": "Arrakis"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Arrakis founder Abhishek Bhardwaj explains why AI agents need secure Linux sandboxes for tool execution, computer use, reinforcement-learning workflows, and recoverable multistep tasks. He introduces Arrakis's Cloud Hypervisor-based microVM isolation, snapshot-and-restore backtracking, browser/VNC access, port forwarding, and Python, Golang, OpenAPI, and MCP interfaces, then examines Linux namespaces, virtualization, root-filesystem isolation, networking, VM construction, and work toward subsecond boot times."
  },
  {
    "slug": "automating-escrow-with-usdc-and-ai",
    "title": "Automating Escrow with USDC and AI",
    "url": "https://ai.engineer/talks/automating-escrow-with-usdc-and-ai",
    "videoId": "AXMdSqdoGHM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 3498000,
    "speakers": [
      {
        "slug": "corey-cooper",
        "name": "Corey Cooper",
        "organization": "Circle"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Circle developer-relations presenter Corey Cooper demonstrates how AI can interpret agreements and verify completed work before Solidity smart contracts release USDC held in escrow. The workshop covers Circle developer tooling, transaction monitoring and webhooks, an open-source escrow-agent demonstration, and audience questions about Ethereum gas costs, document extraction, zero-knowledge proofs, and human approval."
  },
  {
    "slug": "automating-large-scale-refactors-with-parallel-agents",
    "title": "Automating Large-Scale Refactors with Parallel Agents",
    "url": "https://ai.engineer/talks/automating-large-scale-refactors-with-parallel-agents",
    "videoId": "rcsliSIy_YU",
    "event": "AI Engineer Code 2025",
    "durationMs": 4581000,
    "speakers": [
      {
        "slug": "robert-brennan",
        "name": "Robert Brennan",
        "organization": "AllHands"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "OpenHands co-founder Robert Brennan demonstrates how human-supervised teams of parallel coding agents can tackle refactors, technical debt, modernization, and dependency migrations that exceed a single agent’s scope. The workshop covers OpenHands’ origins, Git-based orchestration, a Redux-to-Zustand migration example, practical concurrency limits, the OpenHands Agent and Refactor SDKs, hands-on exercises, and secure Docker-contained agent execution for CVE remediation."
  },
  {
    "slug": "autonomous-agents-for-scientific-tasks-sina-shahandeh-radicait",
    "title": "Autonomous Agents for Scientific Tasks - Sina Shahandeh, RADiCAIT",
    "url": "https://ai.engineer/talks/autonomous-agents-for-scientific-tasks-sina-shahandeh-radicait",
    "videoId": "XLEYtv3cMlw",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1163000,
    "speakers": [
      {
        "slug": "sina-shahandeh",
        "name": "Sina Shahandeh",
        "organization": "RADiCAIT"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Sina Shahandeh explains why autonomous research agents need explicit scientific hypothesis generation rather than merely optimizing code and hyperparameters. Using RADiCAIT's CT-to-synthetic-PET imaging work as a case study, he discusses GAN architectures, 2.5D CT representations, image-fidelity evaluation, Codex-driven experiments, and structured problem decomposition represented in Obsidian. The resulting research loop combines qualitative inspection, hierarchical knowledge, iterative experimentation, and scalable test-time compute."
  },
  {
    "slug": "ax-is-the-only-experience-that-matters",
    "title": "AX is the only Experience that Matters",
    "url": "https://ai.engineer/talks/ax-is-the-only-experience-that-matters",
    "videoId": "e9sLVMN76qU",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 925000,
    "speakers": [
      {
        "slug": "ivan-burazin",
        "name": "Ivan Burazin",
        "organization": "Daytona"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Daytona co-founder Ivan Burazin argues that agent experience should replace human-centered assumptions in developer-tool design. He outlines user-mediated authentication, agent-readable Markdown and llms.txt documentation, and API-first machine interfaces, then describes Daytona's isolated agent-native runtimes, rapid sandbox provisioning, and infrastructure challenges for autonomous code execution."
  },
  {
    "slug": "backlog-md-terminal-kanban-board-for-managing-tasks-with-ai-agents-alex-gavrilescu-funstage",
    "title": "Backlog.md: Terminal Kanban Board for Managing Tasks with AI Agents — Alex Gavrilescu, Funstage",
    "url": "https://ai.engineer/talks/backlog-md-terminal-kanban-board-for-managing-tasks-with-ai-agents-alex-gavrilescu-funstage",
    "videoId": "zMXKhhwiCIc",
    "event": "AI Engineer Code 2025",
    "durationMs": 859000,
    "speakers": [
      {
        "slug": "alex-gavrilescu",
        "name": "Alex Gavrilescu",
        "organization": "Funstage GmbH"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Alex Gavrilescu demonstrates Backlog.md, a terminal Kanban and Markdown-native task manager that helps humans and AI coding agents coordinate through repository-backed tasks. He shows how explicit descriptions, acceptance criteria, implementation plans, and human review checkpoints guide Claude Code, while Model Context Protocol resources and tools expose task workflows to compatible agents including Gemini and Cursor. The presentation concludes that small, auditable tasks improve context management and make failed implementations easier to revise or restart."
  },
  {
    "slug": "balancing-innovation-with-security-safety",
    "title": "Balancing Innovation with Security & Safety",
    "url": "https://ai.engineer/talks/balancing-innovation-with-security-safety",
    "videoId": "G7aSH6N7qY4",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1093000,
    "speakers": [
      {
        "slug": "don-bosco-durai",
        "name": "Don Bosco Durai",
        "organization": "Privacera"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Privacera co-founder and CTO Don Bosco Durai explains how to make AI agents safe for enterprise deployment. He describes risks created by shared-process agent architectures, service credentials, insecure LLMs, and regulated data access; recommends security evaluations, risk scoring, vulnerability and dependency scanning, production monitoring, role-aware authorization, approval thresholds, and human oversight; and introduces the open-source PAIG.ai security and safety project."
  },
  {
    "slug": "bdd-adr-prd-wtf-capturing-decisions-for-humans-and-ai-alike-michal-cichra-safe-intelligence",
    "title": "BDD, ADR, PRD, WTF: Capturing Decisions for Humans and AI Alike — Michal Cichra, Safe Intelligence",
    "url": "https://ai.engineer/talks/bdd-adr-prd-wtf-capturing-decisions-for-humans-and-ai-alike-michal-cichra-safe-intelligence",
    "videoId": "504PvfXou5Y",
    "event": "AI Engineer Europe 2026",
    "durationMs": 769000,
    "speakers": [
      {
        "slug": "michal-cichra",
        "name": "Michal Cichra",
        "organization": "Safe Intelligence"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Safe Intelligence principal engineer Michal Cichra explains how architecture decision records, product requirements documents, and executable Cucumber BDD scenarios preserve intent for both engineers and coding agents. He recommends enforcing architectural boundaries, database isolation, and consistent design systems through import linters, Git hooks, CI, and end-to-end tests instead of depending on prompts or agent memory."
  },
  {
    "slug": "benchmarking-coding-agents-on-new-vs-legacy-codebases-denys-linkov-wisedocs",
    "title": "Benchmarking Coding Agents on New vs Legacy Codebases — Denys Linkov, Wisedocs",
    "url": "https://ai.engineer/talks/benchmarking-coding-agents-on-new-vs-legacy-codebases-denys-linkov-wisedocs",
    "videoId": "7vn4WpqNpck",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1088000,
    "speakers": [
      {
        "slug": "denys-linkov",
        "name": "Denys Linkov",
        "organization": "Wisedocs"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Wisedocs Head of Machine Learning Denys Linkov examines whether a six-month refactor of a legacy, multi-repository medical-claims AI pipeline justified its cost. He describes evaluating five orchestration projects across 17 criteria, prototyping with Temporal, and comparing o3-driven refactoring in Cursor against newer Claude Sonnet 4.6 and Claude Opus 4.8 results. The talk also addresses coding-agent reliability and task-duration limits, and reports lower pipeline costs and processing times, support for larger files, broader developer participation, and feature delivery reduced from months to under a week."
  },
  {
    "slug": "benchmarking-semantic-code-retrieval-on-claude-code",
    "title": "Benchmarking semantic code retrieval on Claude Code",
    "url": "https://ai.engineer/talks/benchmarking-semantic-code-retrieval-on-claude-code",
    "videoId": "zKk7sDMGDEQ",
    "event": "AI Engineer Europe 2026",
    "durationMs": 968000,
    "speakers": [
      {
        "slug": "kuba-rogut",
        "name": "Kuba Rogut",
        "organization": "turbopuffer"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Kuba Rogut of turbopuffer benchmarks semantic code retrieval in Claude Code against conventional grep-based exploration. He explains embeddings as cached computation, demonstrates a TurboGrep/TPuff Search tool that chunks code, embeds it with a Voyage code model, and indexes it in turbopuffer, then compares baseline Claude Code, 50-line windowed reads, and windowed reads augmented with semantic search. The talk contrasts Cursor’s integrated Composer retrieval approach with adding search as an external tool and concludes with audience questions about semantic search, vector databases, and workload tradeoffs."
  },
  {
    "slug": "benchmarks-are-memes-how-what-we-measure-shapes-ai-and-us",
    "title": "Benchmarks Are Memes: How What We Measure Shapes AI—and Us",
    "url": "https://ai.engineer/talks/benchmarks-are-memes-how-what-we-measure-shapes-ai-and-us",
    "videoId": "W3khHzajE04",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 944000,
    "speakers": [
      {
        "slug": "alex-duffy",
        "name": "Alex Duffy",
        "organization": "Every"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Alex Duffy of Every argues that AI benchmarks spread like cultural memes and direct which capabilities models improve. Using benchmark saturation and ChatGPT sycophancy as cautionary examples, he advocates accessible, creative, human-centered evaluations and demonstrates AI Diplomacy, where models including Gemini 2.5 Pro, DeepSeek-R1, and o3 negotiate, form alliances, and reveal contrasting strategic behavior."
  },
  {
    "slug": "benchmarks-the-good-the-bad-and-the-ugly",
    "title": "Benchmarks: The Good, the Bad, and the Ugly",
    "url": "https://ai.engineer/talks/benchmarks-the-good-the-bad-and-the-ugly",
    "videoId": "jWq-aZIU0kM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 769000,
    "speakers": [
      {
        "slug": "ali-khial",
        "name": "Ali Khial",
        "organization": "G2i"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "G2i AI/ML leader Ali Khial examines why software engineering benchmarks can misrepresent real model capability: unrealistic or solution-leaking prompts, economically irrelevant tasks, and weak verifiers that accept incorrect implementations or reject valid alternatives. Drawing on SWE-Bench Pro and DeepSWE examples, he argues that leaderboard distrust reflects a benchmark-quality gap and proposes human-authored, human-reviewed, production-grade tasks grounded in real engineering value."
  },
  {
    "slug": "bending-a-public-mcp-server-without-breaking-it-nimrod-hauser-baz",
    "title": "Bending a Public MCP Server Without Breaking It — Nimrod Hauser, Baz",
    "url": "https://ai.engineer/talks/bending-a-public-mcp-server-without-breaking-it-nimrod-hauser-baz",
    "videoId": "U00AOI1eJUE",
    "event": "AI Engineer Europe 2026",
    "durationMs": 2450000,
    "speakers": [
      {
        "slug": "nimrod-hauser",
        "name": "Nimrod Hauser",
        "organization": "Baz"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Baz founding engineer Nimrod Hauser demonstrates how to adapt a third-party Playwright MCP server for reliable production agents. Using Baz Spec Reviewer, he shows how an agent combines ticket requirements and Figma designs with browser-based implementation checks, then improves integration through tool wrappers, accessibility snapshots, tool filtering and injection, deterministic path-validation guardrails, and practical authentication considerations."
  },
  {
    "slug": "best-practices-for-evaluating-large-language-model-applications-with-llmeval-niklas-nielsen",
    "title": "Best Practices for Evaluating Large Language Model Applications with llmeval: Niklas Nielsen",
    "url": "https://ai.engineer/talks/best-practices-for-evaluating-large-language-model-applications-with-llmeval-niklas-nielsen",
    "videoId": "fiXjTif1nS4",
    "event": "AI Engineer Summit 2023",
    "durationMs": 573000,
    "speakers": [
      {
        "slug": "niklas-nielsen",
        "name": "Niklas Nielsen",
        "organization": "Log10"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Log10 co-founder and CTO Niklas Nielsen demonstrates llmeval, a local Python evaluation framework for making LLM applications more reliable. He covers Hydra-based configuration, prompt and test organization, customizable metrics, strict output comparisons, and reports spanning Claude, GPT-4, and GPT-3.5. The talk also demonstrates model-based grading with explicit criteria, including scored evaluation of Mermaid diagrams and AI-assisted feedback."
  },
  {
    "slug": "beyond-apis-how-ai-web-agents-are-automating-the-long-tail-of-knowledge-work",
    "title": "Beyond APIs: How AI Web Agents Are Automating the \"Long Tail\" of Knowledge Work",
    "url": "https://ai.engineer/talks/beyond-apis-how-ai-web-agents-are-automating-the-long-tail-of-knowledge-work",
    "videoId": "kjSGc7uwDo8",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1064000,
    "speakers": [
      {
        "slug": "arjun-chintapalli",
        "name": "Arjun Chintapalli",
        "organization": "rtrvr.ai"
      },
      {
        "slug": "bhavani-kalisetty",
        "name": "Bhavani Kalisetty",
        "organization": "rtrvr.ai"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Arjun Chintapalli and Bhavani Kalisetty demonstrate rtrvr.ai, also called Retriever, a Chrome-extension-based AI web agent that performs natural-language browser tasks, extracts structured data into Google Sheets, and works across multiple tabs. Examples include LinkedIn interactions, Amazon product comparisons, document summarization, competitive research, and collaborative dataset creation; the presenters contrast their browser-native approach with cloud-hosted browser agents."
  },
  {
    "slug": "beyond-code-coverage-functionality-testing-with-playwright",
    "title": "Beyond Code Coverage: Functionality Testing with Playwright",
    "url": "https://ai.engineer/talks/beyond-code-coverage-functionality-testing-with-playwright",
    "videoId": "FWEInOtngmM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1185000,
    "speakers": [
      {
        "slug": "marlene-mhangami",
        "name": "Marlene Mhangami",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Microsoft and GitHub developer advocate Marlene Mhangami examines whether AI-generated code improves developer productivity, emphasizing that maintainable codebases and meaningful tests matter more than raw output or coverage. She explains red-green test-driven development, critiques excessive reliance on unit tests, and presents a toy-store browser-testing scenario using Playwright MCP and GitHub Copilot. The discussion concludes with practical advice on committing before agent-driven changes, headless execution, and audience questions about Storybook and application state."
  },
  {
    "slug": "beyond-conversation-why-documents-transform-natural-language-into-code",
    "title": "Beyond Conversation: Why Documents Transform Natural Language into Code",
    "url": "https://ai.engineer/talks/beyond-conversation-why-documents-transform-natural-language-into-code",
    "videoId": "2Jom-4Brg6Q",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 657000,
    "speakers": [
      {
        "slug": "filip-kozera",
        "name": "Filip Kozera",
        "organization": "Wordware"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Wordware co-founder and CEO Filip Kozera argues that conversational AI interfaces accumulate polluted context, impede precise iteration, and limit users to one active workflow. He proposes structured documents as clearer specifications for natural-language programming and describes event-triggered background agents operating concurrently under human approval. The talk examines ambient multi-agent coordination, organizational workflows, and the need for communication protocols that express agent constraints and authority beyond current MCP capabilities."
  },
  {
    "slug": "beyond-static-intelligence-evaluating-continual-learning",
    "title": "Beyond Static Intelligence: Evaluating Continual Learning",
    "url": "https://ai.engineer/talks/beyond-static-intelligence-evaluating-continual-learning",
    "videoId": "iqloyWCGYQQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1230000,
    "speakers": [
      {
        "slug": "parth-asawa",
        "name": "Parth Asawa",
        "organization": "UC Berkeley"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "UC Berkeley researcher Parth Asawa argues that conventional independent-task benchmarks measure static capability rather than whether AI agents improve through experience. He presents Continual Learning Bench and explains benchmark headroom, online adaptation, and evaluation across stateful tasks, illustrating learning through repeated SQL queries against unfamiliar databases and adaptation to schema changes. The talk examines memory-management and plasticity failures, balances model strength, learning ability, and computational cost, and closes with implications for open AI research and plans for longer-horizon tasks, open-source models, and parametric approaches."
  },
  {
    "slug": "beyond-the-consensus-navigating-ai-s-frontier-in-2025",
    "title": "Beyond the Consensus: Navigating AI’s Frontier in 2025",
    "url": "https://ai.engineer/talks/beyond-the-consensus-navigating-ai-s-frontier-in-2025",
    "videoId": "HS5a8VIKsvA",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1075000,
    "speakers": [
      {
        "slug": "grace-isford",
        "name": "Grace Isford",
        "organization": "Lux Capital"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Lux Capital partner Grace Isford surveys the 2025 AI frontier, highlighting stronger reasoning models, falling inference costs, open-model competition, and infrastructure investment as catalysts for AI agents. A constrained flight-booking example illustrates how agents mishandle preferences, context, and execution reliability. She recommends stronger system scaffolding, containment of cascading failures, human oversight, self-correcting agents, and differentiated multimodal product experiences."
  },
  {
    "slug": "beyond-the-harness-a-journey-towards-adaptive-engineering-rajiv-chandegra-annicha-labs",
    "title": "Beyond the Harness: A Journey Towards Adaptive Engineering - Rajiv Chandegra, Annicha Labs",
    "url": "https://ai.engineer/talks/beyond-the-harness-a-journey-towards-adaptive-engineering-rajiv-chandegra-annicha-labs",
    "videoId": "qdZzND79mcg",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2221000,
    "speakers": [
      {
        "slug": "rajiv-chandegra",
        "name": "Rajiv Chandegra",
        "organization": "Annicha Labs"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Rajiv Chandegra of Annicha Labs argues that conventional AI-agent harnesses predefine roles, tools, sequencing, and human supervision, making them reliable for bounded software tasks but poorly suited to changing physical and social systems. Drawing on emergence, self-organization, and attractors, he proposes adaptive engineering: designing constraints that allow multi-agent, multi-human, and multi-institutional collaboration to reorganize during runtime. He contrasts established coding and orchestration tools with Hermes Agent's experience-driven adaptation and emphasizes horizontal intelligence for complex real-world problems."
  },
  {
    "slug": "beyond-the-prototype-using-ai-to-write-high-quality-code",
    "title": "Beyond the Prototype: Using AI to Write High-Quality Code",
    "url": "https://ai.engineer/talks/beyond-the-prototype-using-ai-to-write-high-quality-code",
    "videoId": "x_1EumTaXeE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1079000,
    "speakers": [
      {
        "slug": "josh-albrecht",
        "name": "Josh Albrecht",
        "organization": "Imbue"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Imbue CTO Josh Albrecht introduces Sculptor, an experimental coding-agent environment designed to make AI-generated software trustworthy enough for established production codebases. He describes immediate AI-assisted review and four preventive practices—research, planning, specifications, and strict style guides—followed by defect detection using linters, tests, and LLM analysis. Examples include identifying race conditions and exposed API keys, automatically repairing issues, and distinguishing newly introduced lint errors from existing problems."
  },
  {
    "slug": "beyond-transcription-building-voice-ai-that-actually-understands-conversations",
    "title": "Beyond Transcription: Building Voice AI That Actually Understands Conversations",
    "url": "https://ai.engineer/talks/beyond-transcription-building-voice-ai-that-actually-understands-conversations",
    "videoId": "mFLlVpnGpds",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1520000,
    "speakers": [
      {
        "slug": "herve-bredin",
        "name": "Hervé Bredin",
        "organization": "pyannoteAI"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "pyannoteAI co-founder and chief science officer Hervé Bredin explains why speech-to-text alone cannot capture who said what or how conversational context changes meaning. He introduces pyannote speaker diarization, its integration with Whisper and downstream language-model workflows, applications in meeting summaries and podcast analysis, and diarization error rate. A live Python demonstration uses pretrained Hugging Face models with PyTorch and Apple MPS to examine speaker attribution, interruption, and overlapping speech in a recorded telephone conversation."
  },
  {
    "slug": "black-forest-labs-flux-open-research-and-the-future-of-visual-ai",
    "title": "Black Forest Labs: FLUX, Open Research, and the Future of Visual AI",
    "url": "https://ai.engineer/talks/black-forest-labs-flux-open-research-and-the-future-of-visual-ai",
    "videoId": "x8Yb4RidLgM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1352000,
    "speakers": [
      {
        "slug": "stephen-batifol",
        "name": "Stephen Batifol",
        "organization": "Black Forest Labs"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "Black Forest Labs developer advocate Stephen Batifol traces the evolution from FLUX.1 to FLUX.1 Kontext, FLUX.2, and FLUX.2 [klein], highlighting contextual image editing, photorealistic multi-reference generation, and sub-second interactive workflows. He frames these models as steps toward broader visual intelligence, including robotics and action prediction, and presents a SelfFlow robot-manipulation demonstration before audience questions."
  },
  {
    "slug": "blender-mcp-and-the-future-of-creative-tools-siddharth-ahuja",
    "title": "Blender MCP and the Future of Creative Tools - Siddharth Ahuja",
    "url": "https://ai.engineer/talks/blender-mcp-and-the-future-of-creative-tools-siddharth-ahuja",
    "videoId": "nnktgWtfJHE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1010000,
    "speakers": [
      {
        "slug": "siddharth-ahuja",
        "name": "Siddharth Ahuja"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Siddharth Ahuja explains how his Blender MCP project connects AI clients to Blender through the Model Context Protocol and a scripting add-on, allowing natural-language prompts to create and manipulate 3D scenes. He outlines integrations with Sketchfab, Poly Haven, and Rodin, demonstrates generated scenes, reference-image reconstruction, and animation, and considers AI clients as orchestrators across creative applications and external APIs."
  },
  {
    "slug": "bolt-new-how-we-scaled-0-20m-arr-in-60-days-with-15-people",
    "title": "Bolt.new: How we scaled $0-20m ARR in 60 days, with 15 people",
    "url": "https://ai.engineer/talks/bolt-new-how-we-scaled-0-20m-arr-in-60-days-with-15-people",
    "videoId": "s8RM8uYxkoY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1053000,
    "speakers": [
      {
        "slug": "eric-simons",
        "name": "Eric Simons",
        "organization": "Bolt.new / StackBlitz"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Eric Simons describes how Bolt.new emerged after seven years of building StackBlitz and scaled rapidly with a small, tightly aligned team. He emphasizes frugality, focused execution, repeated product experiments, rapid iteration, and transparent public office hours. He also describes using Parahelp's AI support assistant to automate customer-support tickets and increase operational leverage, followed by an audience question about deciding what to build."
  },
  {
    "slug": "books-reimagined-ai-to-create-new-experiences-for-things-you-know-ukasz-gandecki-thebrain-pro",
    "title": "Books reimagined: AI to create new experiences for things you know — Łukasz Gandecki, TheBrain.pro",
    "url": "https://ai.engineer/talks/books-reimagined-ai-to-create-new-experiences-for-things-you-know-ukasz-gandecki-thebrain-pro",
    "videoId": "Kcka7rzcxLk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 584000,
    "speakers": [
      {
        "slug": "ukasz-gandecki",
        "name": "Łukasz Gandecki",
        "organization": "TheBrain.pro"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Łukasz Gandecki demonstrates BookGenius, an AI-enhanced reading experience that adds contextual character information, avatars, music, and embedding-based scene search to books including The Snow Queen and 1984. He explains rapid AI-assisted JavaScript prototyping, selective refactoring, human editorial judgment, and a production pipeline combining LLM-based scene analysis, character detection, mood-aware music generation, and structured XML, concluding with an open-source book-player announcement."
  },
  {
    "slug": "botdojo-launch-enhancing-ai-assistants-with-evaluations-and-synthetic-data",
    "title": "BotDojo Launch: Enhancing AI Assistants with Evaluations and Synthetic Data",
    "url": "https://ai.engineer/talks/botdojo-launch-enhancing-ai-assistants-with-evaluations-and-synthetic-data",
    "videoId": "PNjMwdCo_YM",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 347000,
    "speakers": [
      {
        "slug": "paul-henry",
        "name": "Paul Henry",
        "organization": "BotDojo"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "BotDojo founder Paul Henry demonstrates a low-code, retrieval-backed support chatbot and shows how node tracing, structured outputs, and batch evaluations reveal missing knowledge and hallucination risks. He then generates synthetic documentation from existing indexed material, validates its usefulness, writes it into a vector index, and reruns the evaluation batch, which finishes with improved results."
  },
  {
    "slug": "bounded-autonomy-between-free-will-and-determinism",
    "title": "Bounded Autonomy: Between Free Will and Determinism",
    "url": "https://ai.engineer/talks/bounded-autonomy-between-free-will-and-determinism",
    "videoId": "t4359sKBu4w",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1012000,
    "speakers": [
      {
        "slug": "angus-j-mclean",
        "name": "Angus J. McLean",
        "organization": "OLIVER"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "OLIVER AI director Angus J. McLean argues that effective agentic systems require bounded autonomy rather than unconstrained complexity. Drawing on advertising workflows that he says produce roughly 4,000 creative assets daily for more than 200 brands, he recommends restricting model context, supplying curated documentation instead of unrestricted internet access, embracing creative constraints, and building simple systems with short real-world feedback loops. He closes with an example of clustering 50,000 social posts into usable creative and strategic insights."
  },
  {
    "slug": "break-it-til-you-make-it-building-the-self-improving-stack-for-ai-agents",
    "title": "Break It 'Til You Make It: Building the Self-Improving Stack for AI Agents",
    "url": "https://ai.engineer/talks/break-it-til-you-make-it-building-the-self-improving-stack-for-ai-agents",
    "videoId": "Qvp9vw4jJQ8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 865000,
    "speakers": [
      {
        "slug": "aparna-dhinakaran",
        "name": "Aparna Dhinakaran",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Arize AI co-founder Aparna Dhinakaran explains how to replace ad hoc agent testing with a continuous improvement stack combining production observability, tracing, systematic prompt and model experiments, tool-call and trajectory evaluations, Q&A correctness checks, and multi-turn context assessment. She emphasizes improving evaluators and LLM-as-a-judge prompts alongside the agents themselves, and closes by recommending open-source Arize Phoenix and Arize AX."
  },
  {
    "slug": "breaking-ai-s-1-gigahertz-barrier",
    "title": "Breaking AI’s 1 Gigahertz Barrier",
    "url": "https://ai.engineer/talks/breaking-ai-s-1-gigahertz-barrier",
    "videoId": "gADhNzFjGeI",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1211000,
    "speakers": [
      {
        "slug": "sunny-madra",
        "name": "Sunny Madra",
        "organization": "Groq"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Groq’s Sunny Madra compares rapid AI-inference improvements with the historical race toward gigahertz microprocessors, highlighting a reported greater-than-50% speed increase for Llama 3 8B between April and June. He argues that faster, cheaper token processing can industrialize software and creative work, position LLMs as the core of future computing, and enable new research, multimodal, enterprise, and security applications."
  },
  {
    "slug": "breaking-the-chain-agent-continuations-for-resumable-ai-workflows",
    "title": "Breaking the Chain: Agent Continuations for Resumable AI Workflows",
    "url": "https://ai.engineer/talks/breaking-the-chain-agent-continuations-for-resumable-ai-workflows",
    "videoId": "ZB7l4uxW3Yo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1622000,
    "speakers": [
      {
        "slug": "greg-benson",
        "name": "Greg Benson",
        "organization": "SnapLogic; University of San Francisco"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "SnapLogic chief scientist and University of San Francisco professor Greg Benson introduces Agent Continuations, a programming-language-inspired mechanism for capturing and restoring the complete state of LLM agents, including nested subagents and tool calls. He explains how continuation objects and message histories enable human-approval checkpoints, failure recovery, durable suspension, and resumption without keeping agent loops running. The presentation concludes with a public OpenAI Python API-based prototype and potential integrations with frameworks such as Strands and Pydantic AI."
  },
  {
    "slug": "bringing-agents-onto-the-world-wide-web",
    "title": "Bringing agents onto the world wide web",
    "url": "https://ai.engineer/talks/bringing-agents-onto-the-world-wide-web",
    "videoId": "GqoNrUz8hEU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1106000,
    "speakers": [
      {
        "slug": "paul-klein-iv",
        "name": "Paul Klein IV",
        "organization": "Browserbase"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Browserbase founder Paul Klein IV argues that browser agents are constrained less by model capability than by missing engineering: effective harnesses, appropriate tools, reusable skills and memory, and reliable browser infrastructure. He discusses reinforcement-learning environments, coding-agent harness comparisons, network-request interception, WebMCP and Playwright CLI, and why improvised OpenClaw deployments on home Mac minis do not address enterprise-scale reliability and compliance."
  },
  {
    "slug": "bringing-continual-learning-into-enterprises",
    "title": "Bringing Continual Learning into Enterprises",
    "url": "https://ai.engineer/talks/bringing-continual-learning-into-enterprises",
    "videoId": "ZTA0GwpAUak",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1143000,
    "speakers": [
      {
        "slug": "samuel-denton",
        "name": "Samuel Denton",
        "organization": "Applied Compute"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Applied Compute platform research lead Samuel Denton presents a four-quadrant framework for enterprise continual learning across offline hints, production traces, and on-policy data. Using a Qwen3.5 thinking model on SWE-bench, he describes shortening agent reasoning while preserving test performance, compares offline and on-policy improvements, and discusses reward-shaping regressions, KL learning signals, and relevance-masked self-distillation with an LLM judge."
  },
  {
    "slug": "build-a-prompt-learning-loop",
    "title": "Build a Prompt Learning Loop",
    "url": "https://ai.engineer/talks/build-a-prompt-learning-loop",
    "videoId": "SbcQYbrvAfI",
    "event": "AI Engineer Code 2025",
    "durationMs": 3128000,
    "speakers": [
      {
        "slug": "sally-ann-delucia",
        "name": "SallyAnn DeLucia",
        "organization": "Arize"
      },
      {
        "slug": "fuad-ali",
        "name": "Fuad Ali",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Arize product leaders SallyAnn DeLucia and Fuad Ali present a hands-on workshop on building evaluation-driven prompt-learning loops. They discuss agent failures and planning, meta-prompts that update prompts using scorer feedback, GEPA benchmarking, LLM-as-a-judge and rule-compliance evaluators, iterative metrics, and practical debugging with API keys and nest_asyncio."
  },
  {
    "slug": "build-ai-systems-for-discernment-not-approval-angel-ortmann-lee-duolingo",
    "title": "Build AI Systems for Discernment, Not Approval - Angel Ortmann Lee, Duolingo",
    "url": "https://ai.engineer/talks/build-ai-systems-for-discernment-not-approval-angel-ortmann-lee-duolingo",
    "videoId": "CDqzWpwkSls",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1553000,
    "speakers": [
      {
        "slug": "angel-ortmann-lee",
        "name": "Angel Ortmann Lee",
        "organization": "Duolingo"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Duolingo software engineer Angel Ortmann Lee argues that human oversight of AI is ineffective when interfaces encourage reviewers to rubber-stamp automated judgments. Drawing on Wharton research into cognitive surrender and a published Duolingo English Test experiment involving fabricated cheating alerts, she examines automation bias in high-stakes remote proctoring. She advocates interaction designs that elicit independent reasoning, separate review decisions into clearer components, surface assumptions, and generate more reliable training data."
  },
  {
    "slug": "build-deploy-ai-powered-apps",
    "title": "Build & deploy AI-powered apps",
    "url": "https://ai.engineer/talks/build-deploy-ai-powered-apps",
    "videoId": "G_bHFmEAarM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 3800000,
    "speakers": [
      {
        "slug": "paige-bailey",
        "name": "Paige Bailey",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Google DeepMind developer-relations lead Paige Bailey demonstrates rapid AI application prototyping with Google AI Studio, Gemini, and Gemma. The workshop examines model cost and capability tradeoffs, search and Maps grounding, multimodal embeddings, Firebase-backed applications, Project Genie interactive worlds, generative video and music, open-source robotics, and Gemini integrations with OpenClaw."
  },
  {
    "slug": "build-deploy-ai-powered-apps-18cc1b",
    "title": "Build & deploy AI-powered apps",
    "url": "https://ai.engineer/talks/build-deploy-ai-powered-apps-18cc1b",
    "videoId": "ns9f1fjLD7Y",
    "event": "AI Engineer Europe 2026",
    "durationMs": 6875000,
    "speakers": [
      {
        "slug": "paige-bailey",
        "name": "Paige Bailey",
        "organization": "Google DeepMind"
      },
      {
        "slug": "guillaume-vernade",
        "name": "Guillaume Vernade",
        "organization": "Google DeepMind"
      },
      {
        "slug": "ian-ballantyne",
        "name": "Ian Ballantyne",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Google DeepMind's Paige Bailey demonstrates prompt-driven app creation and an interactive Genie world assembled from generative models. Guillaume Vernade then walks through a book-to-media workflow involving character images, narration, and real-time Lyria music. An additional presenter, Ian Ballantyne, closes with Gemma open models, local and on-device AI, and AI Edge Gallery."
  },
  {
    "slug": "build-dynamic-products-and-stop-the-ai-sideshow",
    "title": "Build Dynamic Products, and Stop the AI Sideshow",
    "url": "https://ai.engineer/talks/build-dynamic-products-and-stop-the-ai-sideshow",
    "videoId": "CB-4NKDYnRs",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1090000,
    "speakers": [
      {
        "slug": "eliza-cabrera",
        "name": "Eliza Cabrera",
        "organization": "Workday"
      },
      {
        "slug": "jeremy-silva",
        "name": "Jeremy Silva",
        "organization": "Freeplay"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Eliza Cabrera of Workday and Jeremy Silva of Freeplay argue that enterprise AI should be integrated into core product strategy rather than isolated in technology-first side projects. They trace the progression from LLM interfaces, multimodal experimentation, RAG, and copilots to tool-using agents, then recommend customer-led discovery, integrated risk planning, aligned teams and roadmaps, and incremental redesign of existing workflows. Workday Help and employee self-service illustrate how AI capabilities can mature within practical enterprise products."
  },
  {
    "slug": "build-enterprise-generative-ai-apps-using-llama-3-at-1-000-tokens-s-on-the-sambanova-ai-platform",
    "title": "Build enterprise generative AI apps using Llama-3 at 1,000 tokens/s on the SambaNova AI platform",
    "url": "https://ai.engineer/talks/build-enterprise-generative-ai-apps-using-llama-3-at-1-000-tokens-s-on-the-sambanova-ai-platform",
    "videoId": "ksgLoPxEQzM",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 3274000,
    "speakers": [
      {
        "slug": "varun-badrinath-krishna",
        "name": "Varun Badrinath Krishna",
        "organization": "SambaNova Systems"
      },
      {
        "slug": "petro-junior-milan",
        "name": "Petro Junior Milan",
        "organization": "SambaNova Systems"
      },
      {
        "slug": "rachelle-mattern",
        "name": "Rachelle Mattern",
        "organization": "SambaNova Systems"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "SambaNova presenters introduce their enterprise AI platform, discuss claimed Llama 3 inference throughput of 1,000 tokens per second, and demonstrate Llama 3 and Samba-1. The hands-on workshop configures Python and the SambaNova API, then builds [REDACTED:password] document-question-answering RAG workflow using document loading, embeddings, semantic retrieval, and ChromaDB vector indexing. The recording also includes an additional self-introduced presenter absent from the two-person published speaker roster."
  },
  {
    "slug": "build-evals-that-actually-matter-nick-ung-akshay-sharma-lyft",
    "title": "Build Evals That Actually Matter - Nick Ung & Akshay Sharma, Lyft",
    "url": "https://ai.engineer/talks/build-evals-that-actually-matter-nick-ung-akshay-sharma-lyft",
    "videoId": "3z2uT5aDx_Y",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2265000,
    "speakers": [
      {
        "slug": "nick-ung",
        "name": "Nick Ung",
        "organization": "Lyft"
      },
      {
        "slug": "akshay-sharma",
        "name": "Akshay Sharma",
        "organization": "Lyft"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Lyft speakers Nick Ung and Akshay Sharma explain how to evaluate multi-turn customer-support AI agents before production deployment. Their approach combines production-representative simulated users, LLM-as-a-judge grading, explicit launch gates, and τ-bench-inspired simulations of LangGraph agents. They emphasize decision-relevant metrics, adequate sample sizes, avoiding evaluation anti-patterns, and continual improvement across the model, context, and evaluation harness."
  },
  {
    "slug": "build-evaluate-and-deploy-a-rag-based-retail-copilot-with-azure-ai",
    "title": "Build, Evaluate and Deploy a RAG-based retail copilot with Azure AI",
    "url": "https://ai.engineer/talks/build-evaluate-and-deploy-a-rag-based-retail-copilot-with-azure-ai",
    "videoId": "T60Tj25J4Zw",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 7078000,
    "speakers": [
      {
        "slug": "cedric-vidal",
        "name": "Cedric Vidal",
        "organization": "Microsoft"
      },
      {
        "slug": "david-smith",
        "name": "David Smith",
        "organization": "Microsoft"
      },
      {
        "slug": "miguel-martinez",
        "name": "Miguel Martinez",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Microsoft advocates David Smith and Cedric Vidal lead a hands-on workshop building, evaluating, and deploying a retail copilot with retrieval-augmented generation, Azure AI Studio, a shared GitHub repository, and GitHub Codespaces. Additional Microsoft technical specialist Miguel Martinez assists participants. The discussion covers vector-backed product retrieval, provisioned Azure resources, application identity with Entra ID, and handling queries for merchandise absent from the retailer's catalog; GPT-4o is mentioned as a model intended for a future iteration of the workshop."
  },
  {
    "slug": "build-for-the-memo-not-the-demo-notes-from-200-investment-committees",
    "title": "Build for the Memo, Not the Demo — Notes from 200 Investment Committees",
    "url": "https://ai.engineer/talks/build-for-the-memo-not-the-demo-notes-from-200-investment-committees",
    "videoId": "tJFjeMBKbIY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1463000,
    "speakers": [
      {
        "slug": "shawn-chan",
        "name": "Shawn Chan",
        "organization": "China Resources Holdings"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Shawn Chan argues that finance AI should produce defensible investment-committee memos rather than impressive demos. Drawing on extensive deal and committee experience, he emphasizes weighting authoritative filings above informal sources, exposing contradictions instead of smoothing them away, linking claims directly to underlying evidence, enforcing consistent figures, and checking founder claims against the data room."
  },
  {
    "slug": "build-systems-not-code",
    "title": "Build Systems, Not Code",
    "url": "https://ai.engineer/talks/build-systems-not-code",
    "videoId": "ZD9-4fW2HhM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1179000,
    "speakers": [
      {
        "slug": "angie-jones",
        "name": "Angie Jones",
        "organization": "Agentic AI Foundation"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Angie Jones argues that AI-assisted software craftsmanship moves upward from writing individual lines of code to architecting dependable agentic systems. Using her Relocation Scout house-hunting agent, she demonstrates workflow design, decomposition of oversized prompts, deterministic automation, structured and queryable agent memory inspired by Karpathy's llm-wiki, state tracking and crash recovery, and maintainable agent documentation through AGENTS.md, skills, scripts, and sub-agents."
  },
  {
    "slug": "build-the-ai-gtm-agent-that-knows-the-buyer-before-the-first-message",
    "title": "Build the AI GTM Agent That Knows the Buyer Before the First Message",
    "url": "https://ai.engineer/talks/build-the-ai-gtm-agent-that-knows-the-buyer-before-the-first-message",
    "videoId": "ltv-L5oMPIs",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1587000,
    "speakers": [
      {
        "slug": "dr-sajjan-kanukolanu",
        "name": "Dr. Sajjan Kanukolanu",
        "organization": "Position²"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Position² executive Sajjan Kanukolanu explains why effective B2B go-to-market agents require coordinated AI, integration, and architecture rather than a chatbot added to existing workflows. He describes combining CRM records, HubSpot conversations, LinkedIn activity, job changes, and website-visitor signals to identify and prioritize buyers, while examining the limits of anonymous identity resolution and feeding sales outcomes back into agent knowledge bases."
  },
  {
    "slug": "build-your-first-demand-driven-context-base-let-ai-agents-tell-you-what-they-need",
    "title": "Build Your First Demand-Driven Context Base: Let AI Agents Tell You What They Need",
    "url": "https://ai.engineer/talks/build-your-first-demand-driven-context-base-let-ai-agents-tell-you-what-they-need",
    "videoId": "_QAVExf_1uw",
    "event": "AI Engineer Europe 2026",
    "durationMs": 4095000,
    "speakers": [
      {
        "slug": "raj-navakoti",
        "name": "Raj Navakoti",
        "organization": "IKEA"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "IKEA staff software engineer Raj Navakoti presents Demand-Driven Context, a methodology for building enterprise AI-agent knowledge bases by starting with real tasks, observing failures, and curating only the missing domain knowledge. The workshop examines agent memory constraints, knowledge fragmented across Confluence, Slack, and GitHub, incident-driven demonstrations, stale documentation, retrieval-layer architecture, and repository-based starter materials. Audience questions explore scalability, expert workload, multi-document retrieval, knowledge maintenance, and protecting proprietary data."
  },
  {
    "slug": "build-your-own-deep-research-agent-technical-writer",
    "title": "Build Your Own Deep Research Agent + Technical Writer",
    "url": "https://ai.engineer/talks/build-your-own-deep-research-agent-technical-writer",
    "videoId": "mYSRn6PC1mc",
    "event": "AI Engineer Europe 2026",
    "durationMs": 7023000,
    "speakers": [
      {
        "slug": "louis-francois-bouchard",
        "name": "Louis-François Bouchard",
        "organization": "Towards AI"
      },
      {
        "slug": "paul-iusztin",
        "name": "Paul Iusztin",
        "organization": "Decoding AI"
      },
      {
        "slug": "samridhi-vaid",
        "name": "Samridhi Vaid",
        "organization": "Towards AI Inc"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Louis-François Bouchard, Paul Iusztin, and Samridhi Vaid demonstrate an end-to-end system combining an MCP-oriented deep research agent with a more constrained technical writing workflow. They explain how Python, uv, GitHub, Gemini, and the Gemini API support grounded research and YouTube analysis; how writing guidelines and structured prompts reduce generic AI-generated content; and how LLM judges, held-out dataset splits, precision, recall, and F1 scores expose quality problems and overfitting."
  },
  {
    "slug": "building-a-10-person-unicorn",
    "title": "Building a 10-Person Unicorn",
    "url": "https://ai.engineer/talks/building-a-10-person-unicorn",
    "videoId": "Qw9P1zvCupE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 723000,
    "speakers": [
      {
        "slug": "max-brodeur-urbas",
        "name": "Max Brodeur-Urbas",
        "organization": "Gumloop"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Gumloop founder Max Brodeur-Urbas explains how his team sought to scale an AI workflow-automation company with fewer than ten people. He describes the transition from an Auto-GPT interface to configurable workflow nodes, selective hiring and recruiting from the customer community, focused team-building sessions, and using Gumloop itself to analyze chatbot conversations and automate internal work."
  },
  {
    "slug": "building-a-chess-coach",
    "title": "Building a Chess Coach",
    "url": "https://ai.engineer/talks/building-a-chess-coach",
    "videoId": "FlzpEGHNVKQ",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1102000,
    "speakers": [
      {
        "slug": "anant-dole",
        "name": "Anant Dole",
        "organization": "Take Take Take"
      },
      {
        "slug": "asbj-rn-steinskog",
        "name": "Asbjørn Steinskog",
        "organization": "Take Take Take"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Anant Dole and Asbjørn Steinskog explain how Take Take Take builds a consumer chess coach by combining Stockfish position analysis and grounded chess signals with language-model-generated explanations. They trace the history of computer chess, examine why general-purpose LLMs struggle to play reliably, demonstrate a Claude-and-Slack feedback loop that revises detectors and regenerates commentary, and compare model accuracy against latency for production use."
  },
  {
    "slug": "building-a-smarter-ai-agent-with-neural-rag",
    "title": "Building a Smarter AI Agent with Neural RAG",
    "url": "https://ai.engineer/talks/building-a-smarter-ai-agent-with-neural-rag",
    "videoId": "xnXqpUW_Kp8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1122000,
    "speakers": [
      {
        "slug": "will-bryk",
        "name": "Will Bryk",
        "organization": "exa"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Exa CEO Will Bryk traces the evolution from Google's PageRank and keyword matching to transformer-based, embedding-driven web search for AI agents. He describes Exa's early model training, why language-model weights cannot contain the entire web, how semantic retrieval handles complex research queries, and a coding demonstration contrasting conventional keyword lookup with an API that produces structured deep-research results."
  },
  {
    "slug": "building-agent-interfaces-lessons-from-chrome-devtools-mcp-for-agents",
    "title": "Building Agent Interfaces: Lessons from Chrome DevTools (MCP) for Agents",
    "url": "https://ai.engineer/talks/building-agent-interfaces-lessons-from-chrome-devtools-mcp-for-agents",
    "videoId": "_B4Pv9ttFgY",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1358000,
    "speakers": [
      {
        "slug": "michael-hablich",
        "name": "Michael Hablich",
        "organization": "Google"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Google Chrome DevTools product manager Michael Hablich explains how to build effective MCP interfaces for coding agents. Using Chrome DevTools for agents and Gemini CLI, he examines semantic summaries of oversized performance traces, the tokens-per-successful-outcome efficiency metric, tool categorization and description trade-offs, actionable error recovery, and maintaining user trust rather than optimizing solely for convenience."
  },
  {
    "slug": "building-agentic-applications-with-heroku-managed-inference-and-agents-julian-duque-and-anush-ds",
    "title": "Building Agentic Applications with Heroku Managed Inference and Agents — Julián Duque and Anush DSouza",
    "url": "https://ai.engineer/talks/building-agentic-applications-with-heroku-managed-inference-and-agents-julian-duque-and-anush-ds",
    "videoId": "vRFqbEzzDsI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 3155000,
    "speakers": [
      {
        "slug": "julian-duque",
        "name": "Julián Duque",
        "organization": "Heroku"
      },
      {
        "slug": "anush-dsouza",
        "name": "Anush DSouza",
        "organization": "Heroku"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Heroku presenters Julián Duque and Anush DSouza lead a hands-on workshop on building agentic applications with Heroku Managed Inference and Agents. Participants deploy a Jupyter Notebook, configure Heroku-hosted tool execution and dyno permissions, inspect database schemas and generate SQL, integrate custom MCP tools including Brave Search, and examine bearer-token-authenticated server-sent-event endpoints."
  },
  {
    "slug": "building-agents-at-cloud-scale-antje-barth-aws",
    "title": "Building Agents at Cloud Scale — Antje Barth, AWS",
    "url": "https://ai.engineer/talks/building-agents-at-cloud-scale-antje-barth-aws",
    "videoId": "WJjInLeaJjo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1140000,
    "speakers": [
      {
        "slug": "antje-barth",
        "name": "Antje Barth",
        "organization": "AWS"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Antje Barth explains how AWS approaches production-scale AI agents, using Alexa Plus and its specialized expert systems to illustrate coordination at scale. She demonstrates Amazon Q Developer CLI integrating MCP servers and grounding responses in AWS documentation, highlights the open-source awslabs/mcp repository, and discusses agent-tool authorization, a D&D-themed demonstration, and related Strands sessions."
  },
  {
    "slug": "building-agents-is-trivial-now-context-is-the-next-frontier",
    "title": "Building agents is trivial now, context is the next frontier",
    "url": "https://ai.engineer/talks/building-agents-is-trivial-now-context-is-the-next-frontier",
    "videoId": "HvMyYLTfvhg",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 802000,
    "speakers": [
      {
        "slug": "jeff-ng",
        "name": "Jeff Ng",
        "organization": "Unblocked"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Unblocked founding engineer Jeff Ng argues that cloud primitives and frameworks such as Flue have simplified production-agent infrastructure, but organizational context remains the primary failure mode. A Linear ticket-enrichment agent recommends re-enabling async dispatch to address QA latency, unaware that the same setting previously caused an outage documented in Slack discussions and a postmortem. Ng contrasts human-supervised local agents with unattended agents, explains why MCP connectivity alone cannot reconcile conflicting information or control context costs, and presents Unblocked's context engine as a way to ground agent plans and code reviews in relevant organizational knowledge."
  },
  {
    "slug": "building-agents-the-hard-parts",
    "title": "Building Agents (the hard parts!)",
    "url": "https://ai.engineer/talks/building-agents-the-hard-parts",
    "videoId": "j_TKDweOsYE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1272000,
    "speakers": [
      {
        "slug": "rita-kozlov",
        "name": "Rita Kozlov",
        "organization": "Cloudflare"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Cloudflare product leader Rita Kozlov explains the shift from AI-assisted work toward autonomous agents and describes four essential building blocks: clients, AI reasoning, workflows, and tools. Through CRM and enterprise-approval examples, the talk examines voice interfaces, MCP integrations, authentication, persistent state with Durable Objects, and Cloudflare’s Agents SDK."
  },
  {
    "slug": "building-agents-with-amazon-nova-act-and-mcp-du-an-lightfoot-and-banjo-obayomi",
    "title": "Building Agents with Amazon Nova Act and MCP — Du’An Lightfoot and Banjo Obayomi",
    "url": "https://ai.engineer/talks/building-agents-with-amazon-nova-act-and-mcp-du-an-lightfoot-and-banjo-obayomi",
    "videoId": "wFTVEDYVJT0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 5180000,
    "speakers": [
      {
        "slug": "du-an-lightfoot",
        "name": "Du’An Lightfoot",
        "organization": "Amazon Web Services"
      },
      {
        "slug": "banjo-obayomi",
        "name": "Banjo Obayomi",
        "organization": "Amazon Web Services"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Du’An Lightfoot and Banjo Obayomi lead a hands-on AWS workshop on building autonomous browser agents with Amazon Nova Act, Model Context Protocol servers, and Strands Agents. They explain AWS agent-building options, provision participant environments, configure Nova Act API access and headless browser execution, discuss integration with Amazon Bedrock, and demonstrate multi-step Google Maps searches and MCP-based PowerPoint generation while answering audience questions."
  },
  {
    "slug": "building-ai-agents-that-actually-automate-knowledge-work",
    "title": "Building AI Agents that actually automate Knowledge Work",
    "url": "https://ai.engineer/talks/building-ai-agents-that-actually-automate-knowledge-work",
    "videoId": "jVGCulhBRZI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1077000,
    "speakers": [
      {
        "slug": "jerry-liu",
        "name": "Jerry Liu",
        "organization": "LlamaIndex"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "LlamaIndex co-founder and CEO Jerry Liu explains how AI agents can automate document-centric knowledge work by moving beyond basic RAG toward document toolboxes that integrate enterprise data sources, permissions, indexing, search, and document manipulation. He discusses combining language and vision models with conventional parsing and agentic validation for complex documents, introduces an Excel agent that normalizes spreadsheets and supports question answering, and outlines human oversight considerations for more autonomous workflows."
  },
  {
    "slug": "building-ai-agents-with-real-roi-in-the-enterprise-sdlc",
    "title": "Building AI Agents with Real ROI in the Enterprise SDLC",
    "url": "https://ai.engineer/talks/building-ai-agents-with-real-roi-in-the-enterprise-sdlc",
    "videoId": "UXOLprPvr-0",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1256000,
    "speakers": [
      {
        "slug": "bruno-passos",
        "name": "Bruno Passos",
        "organization": "booking.com"
      },
      {
        "slug": "beyang-liu",
        "name": "Beyang Liu",
        "organization": "Sourcegraph"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Booking.com's Bruno Passos and Sourcegraph co-founder Beyang Liu explain how their teams apply context-aware coding assistance and software-development agents to enterprise codebases while measuring business impact. Their case study covers developer-experience challenges, experimentation-driven code complexity, Cody and Sourcegraph Search, access to multiple language models, productivity and modernization metrics, feature-flag cleanup, internally developed tools, and automated code review."
  },
  {
    "slug": "building-ai-for-all",
    "title": "Building AI For All",
    "url": "https://ai.engineer/talks/building-ai-for-all",
    "videoId": "ju73sWVtvU0",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1513000,
    "speakers": [
      {
        "slug": "amjad-masad",
        "name": "Amjad Masad",
        "organization": "Replit"
      },
      {
        "slug": "michele-catasta",
        "name": "Michele Catasta",
        "organization": "Replit"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Replit founder Amjad Masad traces programming's progression from punch cards to AI-enhanced development environments and announces broadly available AI coding assistance across Replit. Michele Catasta explains the case for smaller code language models, emphasizing high-quality training data, repeated training passes, HumanEval Pass@1 evaluation, and the release of Replit Code v1.5, concluding with a Perplexity AI collaboration."
  },
  {
    "slug": "building-ai-powered-developer-tools-at-jane-street",
    "title": "Building AI-Powered Developer Tools at Jane Street",
    "url": "https://ai.engineer/talks/building-ai-powered-developer-tools-at-jane-street",
    "videoId": "0ML7ZLMdcl4",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1017000,
    "speakers": [
      {
        "slug": "john-crepezzi",
        "name": "John Crepezzi",
        "organization": "Jane Street"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "John Crepezzi explains how Jane Street builds custom LLM-powered developer tools for its OCaml-centric engineering environment, where scarce public training data, proprietary infrastructure, a Mercurial monorepo, and widespread Emacs usage limit off-the-shelf assistants. He outlines OCaml-based web, editor, and FPGA tooling; domain-specific model training inspired by Meta's CodeCompose; collection of representative developer-workflow data; integrations with VS Code, Emacs, and Neovim; and evaluations that catch code-review-model failures while supporting extensible, pluggable AI infrastructure."
  },
  {
    "slug": "building-ai-products-that-actually-work",
    "title": "Building AI Products That Actually Work",
    "url": "https://ai.engineer/talks/building-ai-products-that-actually-work",
    "videoId": "eSvXbb2EBYc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1122000,
    "speakers": [
      {
        "slug": "ben-hylak",
        "name": "Ben Hylak",
        "organization": "Raindrop"
      },
      {
        "slug": "sid-bendre",
        "name": "Sid Bendre",
        "organization": "Oleve"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Ben Hylak of Raindrop and Sid Bendre of Oleve discuss building dependable AI products through continuous iteration and observation of real-world behavior. Hylak illustrates production failures, questions overreliance on evaluations and language-model judges, and highlights explicit and implicit user-feedback signals. Bendre introduces Oleve’s consumer-product perspective and points attendees to its Trellis framework."
  },
  {
    "slug": "building-alice-s-brain-an-ai-sales-rep-that-learns-like-a-human-sherwood-satwik-11x",
    "title": "Building Alice’s Brain: an AI Sales Rep that Learns Like a Human - Sherwood & Satwik, 11x",
    "url": "https://ai.engineer/talks/building-alice-s-brain-an-ai-sales-rep-that-learns-like-a-human-sherwood-satwik-11x",
    "videoId": "KWmkMV0FNwQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1338000,
    "speakers": [
      {
        "slug": "sherwood-callaway",
        "name": "Sherwood Callaway",
        "organization": "11X"
      },
      {
        "slug": "satwik-singh",
        "name": "Satwik Singh",
        "organization": "11X"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Sherwood Callaway and Satwik Singh explain how 11x equips Alice, its autonomous AI sales development representative, with a customer-specific knowledge base that mimics onboarding a human salesperson. They outline ingestion of sales materials through Amazon S3 and a backend database, document-to-text conversion with LlamaIndex’s LlamaParse, vector embedding and retrieval with Pinecone, evolving RAG techniques, vendor-selection considerations, and future work on detecting hallucinations in generated outreach."
  },
  {
    "slug": "building-an-acp-compatible-agent-live-bennet-fenner-zed",
    "title": "Building an ACP-Compatible Agent Live — Bennet Fenner, Zed",
    "url": "https://ai.engineer/talks/building-an-acp-compatible-agent-live-bennet-fenner-zed",
    "videoId": "HsxQICTLF84",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1099000,
    "speakers": [
      {
        "slug": "bennet-fenner",
        "name": "Bennet Fenner",
        "organization": "Zed"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Zed developer Bennet Fenner explains how Agent Client Protocol standardizes JSON-RPC communication between coding agents and editors, then live-codes a minimal ACP-compatible TypeScript agent backed by the Anthropic API. He covers the agent loop, filesystem tool calls, capability negotiation, session creation and state management, cancellation, Zed integration and debugging, streamed model output, and standard-I/O transport."
  },
  {
    "slug": "building-an-agentic-platform",
    "title": "Building an Agentic Platform",
    "url": "https://ai.engineer/talks/building-an-agentic-platform",
    "videoId": "12v5S1n1eOY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1146000,
    "speakers": [
      {
        "slug": "ben-kus",
        "name": "Ben Kus",
        "organization": "Box"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Box CTO Ben Kus explains how the company evolved from OCR-assisted, single-pass LLM extraction toward an agentic architecture for turning unstructured enterprise documents into structured metadata. He describes accuracy failures on long documents with many complex fields, then outlines coordinated extraction, multiple-model voting, LLM-as-a-judge feedback, and a preference for prompts, prompt caching, and agentic workflows over fine-tuning."
  },
  {
    "slug": "building-an-agentic-video-editor-for-mass-consumer",
    "title": "Building an Agentic Video Editor for Mass Consumer",
    "url": "https://ai.engineer/talks/building-an-agentic-video-editor-for-mass-consumer",
    "videoId": "pPj_tjlvYjA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 765000,
    "speakers": [
      {
        "slug": "ekaterina-deyneka",
        "name": "Ekaterina Deyneka",
        "organization": "Reelful"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Reelful founder and CEO Ekaterina Deyneka explains how a consumer-facing agentic video editor transforms user-provided photos, footage, and lightweight prompts into polished clips. She compares its architecture with agentic app builders: a remote sandbox runs an agent equipped with tools and skills, but produces a video composition and rendered output instead of an application. The talk emphasizes editing real, sometimes incomplete footage, animating photos, using Remotion for composition, and demonstrates an example edited video."
  },
  {
    "slug": "building-an-ai-assistant-that-makes-phone-calls",
    "title": "Building an AI assistant that makes phone calls",
    "url": "https://ai.engineer/talks/building-an-ai-assistant-that-makes-phone-calls",
    "videoId": "idDnpGqJd80",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 3053000,
    "speakers": [
      {
        "slug": "tom-redman",
        "name": "Tom Redman",
        "organization": "Convex"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Convex developer-experience lead Tom Redman demonstrates Floyd, a voice-driven AI assistant that can interpret requests and place telephone calls. The workshop combines Convex, Google Cloud speech-to-text, OpenAI, Twilio, and an Express server, covering reactive database queries, real-time transcript streaming, call-context tracking, generated voice introductions, and audience questions about multimodal models and automated call navigation."
  },
  {
    "slug": "building-an-autonomous-engineering-org",
    "title": "Building an Autonomous Engineering Org",
    "url": "https://ai.engineer/talks/building-an-autonomous-engineering-org",
    "videoId": "whue9_YquGA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1056000,
    "speakers": [
      {
        "slug": "angie-jones",
        "name": "Angie Jones",
        "organization": "Agentic AI Foundation"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Angie Jones describes transforming Block's 3,500-person engineering organization from widespread but low-impact AI-tool usage toward agent-driven software delivery. She explains Block's early work on goose and MCP, introduces a six-stage engineering maturity model, and applies the 1-9-90 rule to cultivate dedicated AI champions. Repository readiness and improved tools ultimately enable employees to delegate bugs and feature implementation to a Slack-based Builder Bot."
  },
  {
    "slug": "building-applications-with-ai-agents",
    "title": "Building Applications with AI Agents",
    "url": "https://ai.engineer/talks/building-applications-with-ai-agents",
    "videoId": "R30col3UPUg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 950000,
    "speakers": [
      {
        "slug": "michael-albada",
        "name": "Michael Albada",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Michael Albada, identifying himself as a Microsoft principal applied scientist working on cybersecurity agents, explains how to build production-ready AI-agent applications around foundation models, carefully scoped tools, adaptable orchestration, and multi-step reasoning. He emphasizes rigorous evaluation and synthetic test generation, recommends tools including IntellAgent, PyRIT, Label Studio, TextGrad, and DSPy, and advocates OpenLLMetry/OpenTelemetry observability, systematic red teaming, and security at every layer."
  },
  {
    "slug": "building-blocks-for-llm-systems-products",
    "title": "Building Blocks for LLM Systems & Products",
    "url": "https://ai.engineer/talks/building-blocks-for-llm-systems-products",
    "videoId": "LzeC1AQ-U5o",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1044000,
    "speakers": [
      {
        "slug": "eugene-yan",
        "name": "Eugene Yan",
        "organization": "Amazon"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Eugene Yan, then a Senior Applied Scientist at Amazon, presents four building blocks for production LLM systems: evaluations, retrieval-augmented generation, guardrails, and user feedback. He explains why generic benchmarks such as MMLU can mislead, recommends small task-specific evaluation sets and deterministic SQL/JSON checks where possible, highlights the expense of LLM-judged open-ended evaluations and the difficulty of document retrieval, and discusses how product UX affects feedback collection."
  },
  {
    "slug": "building-ciso-approved-agent-fleet-architecture",
    "title": "Building CISO-approved agent fleet architecture",
    "url": "https://ai.engineer/talks/building-ciso-approved-agent-fleet-architecture",
    "videoId": "d9rsC6_VLoA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 832000,
    "speakers": [
      {
        "slug": "lou-bichard",
        "name": "Lou Bichard",
        "organization": "Gitpod"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Gitpod Field CTO Lou Bichard explains how the company evolved secure development-environment and AI-agent infrastructure from GCP-hosted, multi-tenant Kubernetes and difficult self-hosted deployments toward a simpler customer-cloud architecture. Its lightweight AWS runner operates as an ECS task, starts isolated development workloads on EC2 with EBS storage, and keeps valuable source code inside customer infrastructure while reducing operational overhead and enterprise security concerns."
  },
  {
    "slug": "building-closed-loop-evals-for-a-multimodal-agent-at-uber-scale",
    "title": "Building Closed-Loop Evals for a Multimodal Agent at Uber Scale",
    "url": "https://ai.engineer/talks/building-closed-loop-evals-for-a-multimodal-agent-at-uber-scale",
    "videoId": "31GUkCBD-Uc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1299000,
    "speakers": [
      {
        "slug": "soumya-gupta",
        "name": "Soumya Gupta",
        "organization": "Uber"
      },
      {
        "slug": "jai-chopra",
        "name": "Jai Chopra",
        "organization": "Uber"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "Soumya Gupta and Jai Chopra explain how Uber evaluates a multimodal agent that enhances Uber Eats merchants’ food photographs while preserving authenticity and customer trust. Their pipeline routes images through understanding, editing, iterative QA, and final publication gates; applies faithfulness checks, bounded regeneration, observability, rollback, and layered safeguards; and monitors marketplace metrics to tune production performance."
  },
  {
    "slug": "building-code-first-ai-agents-with-azure-ai-agent-service-cedric-vidal-microsoft",
    "title": "Building Code-First AI Agents with Azure AI Agent Service — Cédric Vidal, Microsoft",
    "url": "https://ai.engineer/talks/building-code-first-ai-agents-with-azure-ai-agent-service-cedric-vidal-microsoft",
    "videoId": "N4vCBM5YbN0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 6846000,
    "speakers": [
      {
        "slug": "cedric-vidal",
        "name": "Cedric Vidal",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Microsoft principal AI advocate Cédric Vidal leads a hands-on, code-first Python workshop using Azure AI Agent Service to build a sales-oriented conversational agent. The discussion covers specialized-agent routing, MCP, RAG, SQL and PDF data sources, practical environment setup, LLM arithmetic limitations, and evaluating agent tool selection with the Azure AI Evaluation SDK. Participant questions and workshop proctors contribute additional voices."
  },
  {
    "slug": "building-context-aware-reasoning-applications-with-langchain-and-langsmith",
    "title": "Building Context-Aware Reasoning Applications with LangChain and LangSmith",
    "url": "https://ai.engineer/talks/building-context-aware-reasoning-applications-with-langchain-and-langsmith",
    "videoId": "cwjs1WAG9CM",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1134000,
    "speakers": [
      {
        "slug": "harrison-chase",
        "name": "Harrison Chase",
        "organization": "LangChain"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Harrison Chase explains why useful AI assistants require language models embedded in larger context-aware systems, then surveys cognitive architectures ranging from chained model and API calls to model routing and autonomous tool-using agents. He illustrates agent skill development with Voyager and describes how LangChain supports application prototyping while LangSmith provides tracing, debugging, and evaluation visibility. The talk concludes with prompt iteration, LLM-assisted evaluation, production feedback, and collaboration across emerging AI engineering roles."
  },
  {
    "slug": "building-conversational-agents",
    "title": "Building Conversational Agents",
    "url": "https://ai.engineer/talks/building-conversational-agents",
    "videoId": "cVzf49yg0D8",
    "event": "AI Engineer Europe 2026",
    "durationMs": 6454000,
    "speakers": [
      {
        "slug": "thor-schaeff",
        "name": "Thor Schaeff",
        "organization": "Google DeepMind"
      },
      {
        "slug": "philipp-schmid",
        "name": "Philipp Schmid",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Google DeepMind presenters Thor Schaeff and Philipp Schmid lead a hands-on workshop on building conversational agents with Gemini. They contrast generateContent with the Interactions API, demonstrate Python coding-agent tool calls and persistent conversational state, then build real-time voice experiences with the Live API, generated music, configurable system instructions, and the Puck voice. Audience participation and implementation questions address API setup, local development, and personalization."
  },
  {
    "slug": "building-conversational-ai-agents-thor-schaeff-elevenlabs",
    "title": "Building Conversational AI Agents - Thor Schaeff, ElevenLabs",
    "url": "https://ai.engineer/talks/building-conversational-ai-agents-thor-schaeff-elevenlabs",
    "videoId": "MPtCBaZn84A",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 3702000,
    "speakers": [
      {
        "slug": "thor-schaeff",
        "name": "Thor Schaeff",
        "organization": "ElevenLabs"
      },
      {
        "slug": "paul",
        "name": "Paul",
        "organization": "ElevenLabs"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "ElevenLabs developer-experience engineer Thor Schaeff leads a hands-on workshop on multilingual conversational AI agents, demonstrating regional accents, language-specific text-to-speech voices, and the Voice Library. Attendee questions explore agent model selection, latency and RAG, multilingual transcription limits, voice-creator moderation, and pronunciation of specialized acronyms; a colleague identified as Paul assists the workshop."
  },
  {
    "slug": "building-cursor-composer",
    "title": "Building Cursor Composer",
    "url": "https://ai.engineer/talks/building-cursor-composer",
    "videoId": "fL1iJHtl51Q",
    "event": "AI Engineer Code 2025",
    "durationMs": 936000,
    "speakers": [
      {
        "slug": "lee-robinson",
        "name": "Lee Robinson",
        "organization": "Cursor"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Lee Robinson explains how Cursor built Composer, its low-latency agentic coding model, balancing coding quality with faster token generation. The talk covers realistic internal evaluations, reinforcement-learning rollouts, mixture-of-experts training across thousands of GPUs, PyTorch and Ray infrastructure, optimized NVIDIA Blackwell kernels, and reuse of cloud-agent virtual-machine fleets to align sandboxed training environments with production. Robinson concludes that reinforcement learning improves both response speed and effective agent behavior."
  },
  {
    "slug": "building-deterministic-infrastructure-for-non-deterministic-ai-agents",
    "title": "Building Deterministic Infrastructure for Non-Deterministic AI Agents",
    "url": "https://ai.engineer/talks/building-deterministic-infrastructure-for-non-deterministic-ai-agents",
    "videoId": "APh1Vx0oLmQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 434000,
    "speakers": [
      {
        "slug": "nishant-gupta",
        "name": "Nishant Gupta",
        "organization": "Meta"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Meta engineering lead Nishant Gupta argues that autonomous AI agents must be operated as distributed systems whose deterministic infrastructure compensates for stochastic model behavior. He examines long-running stateful workflows, retry amplification and runaway GPU costs, shared-memory consistency, agentic control planes, observability, human oversight, and elastic workload scheduling."
  },
  {
    "slug": "building-durable-agents-with-workflow-devkit-ai-sdk",
    "title": "Building durable Agents with Workflow DevKit & AI SDK",
    "url": "https://ai.engineer/talks/building-durable-agents-with-workflow-devkit-ai-sdk",
    "videoId": "kmV-qg4uoNI",
    "event": "AI Engineer Code 2025",
    "durationMs": 4189000,
    "speakers": [
      {
        "slug": "peter-wielander",
        "name": "Peter Wielander",
        "organization": "Vercel"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Vercel’s Peter Wielander leads an in-person, hands-on workshop converting the open-source Vibe Coding Platform coding agent into a durable, observable workflow using Workflow DevKit and AI SDK. He demonstrates repository-based checkpoints, workflow and step directives, agent tool calls, deployment portability, resumable streams, webhooks, suspended and scheduled execution, human-in-the-loop patterns, and operational observability, with extensive audience questions."
  },
  {
    "slug": "building-durable-production-ready-agents-with-openai-sdk-and-temporal",
    "title": "Building Durable, Production-Ready Agents with OpenAI SDK and Temporal",
    "url": "https://ai.engineer/talks/building-durable-production-ready-agents-with-openai-sdk-and-temporal",
    "videoId": "k8cnVCMYmNc",
    "event": "AI Engineer Code 2025",
    "durationMs": 4710000,
    "speakers": [
      {
        "slug": "cornelia-davis",
        "name": "Cornelia Davis",
        "organization": "Temporal"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Temporal developer advocate Cornelia Davis demonstrates how integrating the OpenAI Agents SDK with Temporal makes AI-agent workflows durable and production-ready. The workshop covers persisted LLM calls that avoid duplicate token spending, state management, event sourcing and worker recovery, activity queues, AI Cookbook agentic-loop examples, the Temporal UI, and exposing workflow activities as agent tools through activity_as_tool, followed by operational audience questions."
  },
  {
    "slug": "building-effective-voice-agents",
    "title": "Building Effective Voice Agents",
    "url": "https://ai.engineer/talks/building-effective-voice-agents",
    "videoId": "-OXiljTJxQU",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1037000,
    "speakers": [
      {
        "slug": "anoop-kotha",
        "name": "Anoop Kotha",
        "organization": "OpenAI"
      },
      {
        "slug": "toki-sherbakov",
        "name": "Toki Sherbakov",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "OpenAI solution architects Toki Sherbakov and Anoop Kotha explain how to build production voice agents, contrasting chained transcription–LLM–text-to-speech pipelines with low-latency speech-to-speech applications using the Realtime API. They discuss customer-service trade-offs involving accuracy, determinism, latency, and telephony integrations, then cover delegation to models such as o3 and o4-mini, constrained tool use, prompting, Agents SDK handoffs, evaluations, and observability."
  },
  {
    "slug": "building-efficient-hybrid-context-query-for-llm-grounding",
    "title": "Building efficient hybrid context query for LLM grounding",
    "url": "https://ai.engineer/talks/building-efficient-hybrid-context-query-for-llm-grounding",
    "videoId": "pijYURicI1Y",
    "event": "AI Engineer Summit 2023",
    "durationMs": 695000,
    "speakers": [
      {
        "slug": "simrat-hanspal",
        "name": "Simrat Hanspal",
        "organization": "Hasura"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Hasura technical evangelist Simrat Hanspal demonstrates a secure hybrid RAG pipeline for e-commerce product search. The talk combines semantic vector retrieval and structured filters through a unified Hasura GraphQL API, demonstrates Weaviate, Postgres integration, automatic vectorization, a Streamlit interface, and OpenAI-powered query generation, and concludes by restricting application permissions to prevent unintended operations."
  },
  {
    "slug": "building-enterprise-llm-agents-that-work",
    "title": "Building enterprise LLM agents that work",
    "url": "https://ai.engineer/talks/building-enterprise-llm-agents-that-work",
    "videoId": "4J8-D0sgU9A",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1109000,
    "speakers": [
      {
        "slug": "shaan-desai",
        "name": "Shaan Desai",
        "organization": "Cohere"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Cohere machine learning engineer Shaan Desai explains how to build reliable enterprise LLM agents, comparing LangGraph, LangChain, AutoGen, and CrewAI and outlining when to use single-agent or multi-agent architectures. He discusses precise tool specifications, conversation-history caching, human approval for sensitive actions, observability, golden-set evaluation, and failure mitigation through prompt engineering, synthetic data, and fine-tuning. The talk concludes with BFCL V3 tool-calling evaluation and Cohere's North enterprise platform."
  },
  {
    "slug": "building-generative-image-video-models-at-scale",
    "title": "Building Generative Image & Video Models at Scale",
    "url": "https://ai.engineer/talks/building-generative-image-video-models-at-scale",
    "videoId": "xOP1PM8fwnk",
    "event": "AI Engineer Europe 2026",
    "durationMs": 2446000,
    "speakers": [
      {
        "slug": "sander-dieleman",
        "name": "Sander Dieleman",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "Google DeepMind research scientist Sander Dieleman explains how large-scale generative image and video systems are built, using Veo and Nano Banana as examples. He covers latent representations and Stable Diffusion-style autoencoders; diffusion versus autoregression; noise, frequency spectra, and scaling behavior; U-Net architectures and joint video denoising; sampling guidance and GLIDE; and conditioning signals introduced during post-training. Audience questions address model style, text diffusion, and JAX versus PyTorch for sharding and parallelism."
  },
  {
    "slug": "building-in-the-gemini-era-kat-kampf-ammaar-reshi-google-deepmind",
    "title": "Building in the Gemini Era – Kat Kampf & Ammaar Reshi, Google DeepMind",
    "url": "https://ai.engineer/talks/building-in-the-gemini-era-kat-kampf-ammaar-reshi-google-deepmind",
    "videoId": "fgkXEIbZpGc",
    "event": "AI Engineer Code 2025",
    "durationMs": 1077000,
    "speakers": [
      {
        "slug": "kat-kampf",
        "name": "Kat Kampf",
        "organization": "Google DeepMind"
      },
      {
        "slug": "ammaar-reshi",
        "name": "Ammaar Reshi",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "Google DeepMind product manager Kat Kampf and Google AI Studio product-and-design lead Ammaar Reshi demonstrate Gemini 3 Pro and Nano Banana Pro, Google's Gemini 3 Pro Image model. They show how Google AI Studio turns natural-language prompts into applications, integrates the Gemini API, Live API, and Google Search and Maps grounding, and supports interactive image-driven storytelling and a live multiplayer game demonstration. The presentation emphasizes stronger interface generation, richer image controls and world knowledge, and broader access to software creation."
  },
  {
    "slug": "building-intelligent-research-agents-with-manus",
    "title": "Building Intelligent Research Agents with Manus",
    "url": "https://ai.engineer/talks/building-intelligent-research-agents-with-manus",
    "videoId": "xz0-brt56L8",
    "event": "AI Engineer Code 2025",
    "durationMs": 4890000,
    "speakers": [
      {
        "slug": "ivan-leo",
        "name": "Ivan Leo",
        "organization": "Manus"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Ivan Leo demonstrates Manus 1.5 and the Manus API in a hands-on workshop on building research and workflow agents. He introduces Manus applications across Slack, browsers, email, and generated web apps, then works through asynchronous tasks, polling, file uploads, webhook processing with Modal, and a Slack bot that maintains task context across conversation threads. The session ends with audience questions and practical guidance about Manus integrations."
  },
  {
    "slug": "building-interactive-uis-in-vs-code-with-mcp-apps-marlene-mhangami-liam-hampton-github",
    "title": "Building Interactive UIs in VS Code with MCP Apps — Marlene Mhangami & Liam Hampton, GitHub",
    "url": "https://ai.engineer/talks/building-interactive-uis-in-vs-code-with-mcp-apps-marlene-mhangami-liam-hampton-github",
    "videoId": "_xIwFcnHqp4",
    "event": "AI Engineer Europe 2026",
    "durationMs": 966030,
    "speakers": [
      {
        "slug": "marlene-mhangami",
        "name": "Marlene Mhangami",
        "organization": "Microsoft"
      },
      {
        "slug": "liam-hampton",
        "name": "Liam Hampton",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Marlene Mhangami and Liam Hampton explain how MCP Apps extend Model Context Protocol tools with interactive interfaces rendered directly inside VS Code chat. They outline MCP hosts, clients, and servers; explain UI-resource fetching and sandboxed iframe isolation; and discuss Excalidraw diagrams and Shopify commerce experiences. A live demonstration uses an MCP repository skill with GitHub Copilot CLI to build a Go profiling interface that visualizes bubble-sort performance as an interactive flame graph."
  },
  {
    "slug": "building-metrics-that-actually-work-david-karam-pi-labs",
    "title": "Building Metrics That Actually Work — David Karam, Pi Labs",
    "url": "https://ai.engineer/talks/building-metrics-that-actually-work-david-karam-pi-labs",
    "videoId": "jxrGodnopHo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2428000,
    "speakers": [
      {
        "slug": "david-karam",
        "name": "David Karam",
        "organization": "Pi Labs"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "A hands-on Pi Labs workshop applies lessons from Google Search to evaluating stochastic LLM applications. The presenters discuss defining application-specific metrics, combining natural-language questions with generated Python checks, creating synthetic examples, calibrating scores against data, and using an evaluation copilot, spreadsheets, an SDK, and Google Colab to build practical scoring workflows."
  },
  {
    "slug": "building-multi-agent-systems-with-finite-state-machines",
    "title": "Building Multi-agent Systems with Finite State Machines",
    "url": "https://ai.engineer/talks/building-multi-agent-systems-with-finite-state-machines",
    "videoId": "OD13PiXw60o",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1031000,
    "speakers": [
      {
        "slug": "adam-terlson",
        "name": "Adam Terlson",
        "organization": "Best Buy"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Adam Terlson of Best Buy explains how finite state machines and the Actor model add predictable behavior, observability, recoverability, and operational controls to LLM-powered multi-agent systems. He illustrates ordered messaging with Azure Service Bus sessions and Amazon SQS FIFO message groups, then develops patterns for tool calling, retrieval-augmented generation, human approval, feedback, collaboration, guarded orchestration, and LLM-generated statecharts."
  },
  {
    "slug": "building-multimodal-ai-agents-from-scratch",
    "title": "Building Multimodal AI Agents (From Scratch)",
    "url": "https://ai.engineer/talks/building-multimodal-ai-agents-from-scratch",
    "videoId": "640KMYtxCeI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2218000,
    "speakers": [
      {
        "slug": "apoorva-joshi",
        "name": "Apoorva Joshi",
        "organization": "MongoDB"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "MongoDB developer advocate Apoorva Joshi introduces a hands-on Python workshop for building multimodal AI agents. She contrasts prompting, RAG, and agents; explains agent reasoning, perception, memory, text-image inputs, and application-executed functions; and describes using Voyage AI multimodal embeddings with MongoDB as a vector database. The recording ends as participants begin an approximately 45-minute hands-on lab."
  },
  {
    "slug": "building-protected-mcp-servers",
    "title": "Building Protected MCP Servers",
    "url": "https://ai.engineer/talks/building-protected-mcp-servers",
    "videoId": "PHBGhUKAM-w",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1213000,
    "speakers": [
      {
        "slug": "den-delimarsky-devdiv",
        "name": "Den Delimarsky (DEVDIV)"
      },
      {
        "slug": "julia-kasper",
        "name": "Julia Kasper",
        "organization": "Microsoft"
      },
      {
        "slug": "den-delimarsky",
        "name": "Den Delimarsky",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Microsoft engineers Den Delimarsky and Julia Kasper explain why internet-accessible MCP servers must preserve user identity and authorization when invoking protected APIs. They contrast the March 2025 authorization model with an emerging design that separates MCP resource servers from OAuth authorization servers, delegates token issuance to identity providers, and discovers authorization services through metadata. Demonstrations cover server authentication configuration, token validation, and accessing protected MCP tools from VS Code."
  },
  {
    "slug": "building-reactive-ai-apps",
    "title": "Building Reactive AI Apps",
    "url": "https://ai.engineer/talks/building-reactive-ai-apps",
    "videoId": "qpPgCA664xw",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1022000,
    "speakers": [
      {
        "slug": "matt-welsh",
        "name": "Matt Welsh",
        "organization": "Fixie.ai"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Fixie co-founder Matt Welsh introduces AI.JSX, an open-source TypeScript/JSX framework for building reactive, composable LLM applications. He explains how it simplifies infrastructure concerns such as vector databases, context windows, RAG, and tools; supports Anthropic and OpenAI; demonstrates component-based output filtering and GitHub GraphQL integration; and concludes with a low-latency conversational voice-agent demonstration."
  },
  {
    "slug": "building-redacted-username-gergely-orosz-simon-eskildsen",
    "title": "Building: Gergely Orosz × Simon Eskildsen",
    "url": "https://ai.engineer/talks/building-redacted-username-gergely-orosz-simon-eskildsen",
    "videoId": "jQDXzEVHMSE",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 3390000,
    "speakers": [
      {
        "slug": "gergely-orosz",
        "name": "Gergely Orosz",
        "organization": "The Pragmatic Engineer"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Gergely Orosz interviews [REDACTED:username] co-founder and CEO Simon Eskildsen about discovering programming through early Microsoft and web-development tools, scaling infrastructure at Shopify with future co-founder Justine Li, and building a pragmatic search database around customer needs and infrastructure economics. Their discussion also covers recommendation systems, database and caching design, Cursor and cloud costs, fundraising and employee dilution, and [REDACTED:username]'s distributed-team culture."
  },
  {
    "slug": "building-reliable-agentic-systems",
    "title": "Building Reliable Agentic Systems",
    "url": "https://ai.engineer/talks/building-reliable-agentic-systems",
    "videoId": "bjNYEc908oQ",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1094000,
    "speakers": [
      {
        "slug": "eno-reyes",
        "name": "Eno Reyes",
        "organization": "Factory"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Factory speaker Eno Reyes explains lessons from building Droids, autonomous agents for code review, documentation, testing, and software-development tasks. He discusses agent planning and sub-task decomposition, explicit decision criteria, inference-cost trade-offs, simulated decision paths and Language Agent Tree Search, environmental grounding through custom tool interfaces, long-context models, and the balance between autonomy and human oversight."
  },
  {
    "slug": "building-reliable-support-agents-using-the-effect-typescript-library-michael-fester",
    "title": "Building Reliable Support Agents Using the Effect TypeScript Library - Michael Fester",
    "url": "https://ai.engineer/talks/building-reliable-support-agents-using-the-effect-typescript-library-michael-fester",
    "videoId": "sXXl3YMU7ZI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 442000,
    "speakers": [
      {
        "slug": "michael-fester",
        "name": "Michael Fester",
        "organization": "14.ai"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "14.ai co-founder and CTO Michael Fester explains how his team builds reliable AI-native customer support agents using Effect and TypeScript. He describes a type-safe architecture spanning React, Effect RPC and HTTP, generated OpenAPI documentation, Postgres, and Effect SQL; a composable workflow DSL; provider failover between GPT-4o mini and Gemini 2.0 Flash; stateful retries, token-stream duplication, OpenTelemetry observability, and dependency-injection-based testing. He also discusses Effect's learning curve and the difficulty of tracing dependency injection across larger systems."
  },
  {
    "slug": "building-safe-payment-infrastructure-for-the-autonomous-economy",
    "title": "Building safe Payment Infrastructure for the autonomous economy",
    "url": "https://ai.engineer/talks/building-safe-payment-infrastructure-for-the-autonomous-economy",
    "videoId": "KLSuFPj2ld0",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1126000,
    "speakers": [
      {
        "slug": "steve-kaliski",
        "name": "Steve Kaliski",
        "organization": "Stripe"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Stripe principal software engineer Steve Kaliski explains how autonomous agents can transact safely while separating nondeterministic discovery from deterministic credentials, checkout, and payments. He examines domain-verification and credential risks, demonstrates spending controls and shared payment tokens integrated with Stripe PaymentIntents, and introduces the Stripe-and-Tempo Machine Payments Protocol for paid HTTP requests using HTTP 402. A demonstration includes the Tempo blockchain and pathUSD, followed by audience questions about recurring agent budgets and transaction volumes."
  },
  {
    "slug": "building-scaling-an-ai-agent-swarm-of-low-latency-real-time-voice-bots",
    "title": "Building & Scaling an AI Agent Swarm of low latency real time voice bots!",
    "url": "https://ai.engineer/talks/building-scaling-an-ai-agent-swarm-of-low-latency-real-time-voice-bots",
    "videoId": "7_WhRAuP2Wg",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 4043000,
    "speakers": [
      {
        "slug": "damien-murphy",
        "name": "Damien Murphy",
        "organization": "Deepgram"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Deepgram senior applied engineer Damien Murphy leads a hands-on workshop on building and scaling low-latency conversational voice agents. He demonstrates a Node.js and browser-based JavaScript setup using public client/server repositories, then discusses API access, audio bandwidth, function calling, agent routing, Twilio-style telephony deployment, monitoring, scaling, and speaker-identification tradeoffs."
  },
  {
    "slug": "building-security-around-ml",
    "title": "Building security around ML",
    "url": "https://ai.engineer/talks/building-security-around-ml",
    "videoId": "GVbPiq3Pet0",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1501000,
    "speakers": [
      {
        "slug": "dr-andrew-davis",
        "name": "Dr. Andrew Davis",
        "organization": "HiddenLayer"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "HiddenLayer chief data scientist Dr. Andrew Davis outlines practical defenses for machine-learning and multimodal AI systems, covering adversarial examples, production API monitoring, software vulnerabilities, and data poisoning. He uses ImageNet's externally hosted images to illustrate expired-domain and provenance risks, recommends checksum verification and skeptical handling of VirusTotal and RAG sources, examines Stanford Alpaca's low-cost replication of model behavior, and discusses encoding-related attack surfaces and prompt injection against email-connected agents."
  },
  {
    "slug": "building-self-coding-agents",
    "title": "Building Self-Coding Agents",
    "url": "https://ai.engineer/talks/building-self-coding-agents",
    "videoId": "Iw_3cRf3lnM",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1043000,
    "speakers": [
      {
        "slug": "colin-flaherty",
        "name": "Colin Flaherty",
        "organization": "Augment Code"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Augment Code founding researcher Colin Flaherty explains how his team built an AI coding agent that helped extend its own codebase, adding integrations, searching unfamiliar API documentation, and writing tests. He describes planning agents, tool use, memory, recursive execution, enterprise-grade codebase context, and human-agent interface design, while examining practical limitations through Graphite workflows and difficult parallel-programming errors."
  },
  {
    "slug": "building-sota-open-weights-tool-use-the-command-r-family",
    "title": "Building SOTA Open Weights Tool Use: The Command R Family",
    "url": "https://ai.engineer/talks/building-sota-open-weights-tool-use-the-command-r-family",
    "videoId": "vNssL4u5jb8",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 903000,
    "speakers": [
      {
        "slug": "sandra-kublik",
        "name": "Sandra Kublik",
        "organization": "Cohere"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Cohere developer advocate Sandra Kublik introduces the open-weight Command R and Command R+ model family, emphasizing scalable retrieval-augmented generation, advanced reasoning, and enterprise multi-step tool use. The presentation discusses community adoption through Hugging Face, HuggingChat, and OpenRouter; practical RAG challenges including prompt sensitivity and document-position bias; KILT benchmarking; and an application toolkit using Next.js and SQL."
  },
  {
    "slug": "building-the-platform-for-agent-coordination",
    "title": "Building the platform for agent coordination",
    "url": "https://ai.engineer/talks/building-the-platform-for-agent-coordination",
    "videoId": "UG9IAdmi2Dg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1183000,
    "speakers": [
      {
        "slug": "tom-moor",
        "name": "Tom Moor",
        "organization": "Linear"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Linear Head of Engineering Tom Moor describes the company's progression from pragmatic, embedded AI features toward a platform that coordinates human and AI teammates. He explains its initial use of OpenAI embeddings with pgvector on GCP, similar-issue detection and Slack-based issue creation, a later move to Cohere embeddings, and an architecture where agents have identities and participate as first-class workspace users. He closes with practical guidance for agent behavior: communicate plans, clarify intent, and act like a useful teammate."
  },
  {
    "slug": "building-trust-in-enterprise-ai-evaluating-domain-specific-llms-for-real-world-financial-scenari",
    "title": "Building Trust in Enterprise AI: Evaluating Domain-Specific LLMs for Real-World Financial Scenarios",
    "url": "https://ai.engineer/talks/building-trust-in-enterprise-ai-evaluating-domain-specific-llms-for-real-world-financial-scenari",
    "videoId": "pPvoLjYj_mY",
    "event": "AI Engineer Summit 2025",
    "durationMs": 721000,
    "speakers": [
      {
        "slug": "waseem-alshikh",
        "name": "Waseem Alshikh",
        "organization": "Writer"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Writer co-founder and CTO Waseem Alshikh explains why strong general-purpose benchmark accuracy does not guarantee reliable financial AI. He introduces FailSafeQA, a financial long-context evaluation covering misspelled, incomplete, and out-of-domain queries alongside missing, OCR-corrupted, and irrelevant context. Comparing reasoning models with finance-specialized Palmyra models, he argues that trustworthy enterprise systems must balance answer robustness with context grounding and decline to answer when supporting information is inadequate."
  },
  {
    "slug": "building-voice-agents-with-openai",
    "title": "Building voice agents with OpenAI",
    "url": "https://ai.engineer/talks/building-voice-agents-with-openai",
    "videoId": "iXhba366fQc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 5135000,
    "speakers": [
      {
        "slug": "dominik-kundel",
        "name": "Dominik Kundel",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "OpenAI developer-experience leader Dominik Kundel leads a hands-on World's Fair workshop on building browser-based voice agents with the OpenAI Agents SDK for TypeScript. He introduces native realtime voice support, develops a Next.js application from an existing text-agent foundation, and demonstrates tool calls, human approval, interruption handling, conversation context, agent handoffs, and transcript-based output guardrails. Audience questions address Lemonade dashboards, PCM16 audio formats, and migration from conventional agents to realtime agents."
  },
  {
    "slug": "building-with-anthropic-s-claude-the-prompt-doctor-is-in",
    "title": "Building with Anthropic's Claude - The Prompt Doctor is In",
    "url": "https://ai.engineer/talks/building-with-anthropic-s-claude-the-prompt-doctor-is-in",
    "videoId": "hkhDdcM5V94",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 5696000,
    "speakers": [
      {
        "slug": "jamie-neuwirth",
        "name": "Jamie Neuwirth",
        "organization": "Anthropic"
      },
      {
        "slug": "zack-witten",
        "name": "Zack Witten",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Anthropic's Jamie Neuwirth introduces startup support, rate limits, and Claude 3.5 Sonnet before Zack Witten leads an interactive workshop testing and revising audience-submitted prompts in the Anthropic Console. Examples address concise conversational agents, missing context, structured JSON and code-editing outputs, unit tests, handwriting and OCR, Japanese tokenization, and API log probabilities."
  },
  {
    "slug": "building-your-own-secure-ai-workflows-human-in-the-loop-automation-with-n8n",
    "title": "Building Your Own Secure AI Workflows: Human-in-the-Loop Automation with n8n",
    "url": "https://ai.engineer/talks/building-your-own-secure-ai-workflows-human-in-the-loop-automation-with-n8n",
    "videoId": "tDArkCqjA-c",
    "event": "AI Engineer Europe 2026",
    "durationMs": 4762000,
    "speakers": [
      {
        "slug": "liam-mcgarrigle",
        "name": "Liam McGarrigle",
        "organization": "n8n"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "n8n developer advocate Liam McGarrigle leads a hands-on workshop building a visual AI agent for Gmail and Google Calendar management. He demonstrates conditional integrations, conversation-memory limits, Gmail execution logs and human approval, then discusses MCP workflow safeguards, collaborative governance, and specialized agents using different LLMs."
  },
  {
    "slug": "building-your-own-software-factory",
    "title": "Building your own software factory",
    "url": "https://ai.engineer/talks/building-your-own-software-factory",
    "videoId": "rnDm57Py54A",
    "event": "AI Engineer Europe 2026",
    "durationMs": 5017000,
    "speakers": [
      {
        "slug": "eric-zakariasson",
        "name": "Eric Zakariasson",
        "organization": "Cursor"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Cursor engineer Eric Zakariasson explains how teams can move beyond AI pair programming toward a software factory built around progressively more autonomous coding agents. Using Dan Shapiro's autonomy framework, he discusses agent-first workflows, scheduled MCP-enabled automation, managing multiple and nested agents, and retaining human oversight through observability, specifications, evolving rules, and architectural guardrails. Audience questions probe code quality, enterprise brownfield adoption, cloud-agent costs, and local execution using development containers."
  },
  {
    "slug": "buy-now-maybe-pay-later-dealing-with-prompt-tax-while-staying-at-the-frontier-andrew-thompson",
    "title": "Buy Now, Maybe Pay Later: Dealing with Prompt-Tax While Staying at the Frontier - Andrew Thompson",
    "url": "https://ai.engineer/talks/buy-now-maybe-pay-later-dealing-with-prompt-tax-while-staying-at-the-frontier-andrew-thompson",
    "videoId": "Bf71xMwd-Y0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1509000,
    "speakers": [
      {
        "slug": "andrew-thompson",
        "name": "Andrew Thompson",
        "organization": "Orbital"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Orbital CTO Andrew Thompson explains the “prompt tax”: the engineering and domain-expert effort required when rapidly improving frontier models break existing prompts or introduce regressions. Using Orbital Copilot, an agentic real-estate legal assistant, he demonstrates OCR-driven document processing, task decomposition, lawyer-reviewed outputs, migrations across GPT and reasoning models, and cross-functional prompt ownership. He discusses rigorous experimentation, anticipating future model capabilities, progressive rollouts, and the prospective role of evaluations in managing migration risk."
  },
  {
    "slug": "bypassing-the-multimodal-tax-framework-free-hybrid-rag-raw-sql-rrf-and-live-ui-telemetry",
    "title": "Bypassing the Multimodal Tax: Framework-Free Hybrid RAG, Raw SQL RRF, and Live UI Telemetry",
    "url": "https://ai.engineer/talks/bypassing-the-multimodal-tax-framework-free-hybrid-rag-raw-sql-rrf-and-live-ui-telemetry",
    "videoId": "Akm1sqvWG4A",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2748000,
    "speakers": [
      {
        "slug": "abed-matini",
        "name": "Abed Matini",
        "organization": "Ogilvy"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Ogilvy senior backend developer Abed Matini demonstrates a local-first employee-handbook FAQ assistant and explains framework-light hybrid RAG using document ingestion, chunking, PostgreSQL retrieval, and Reciprocal Rank Fusion. He outlines a Python/FastAPI, React, Docker, Ollama, and GitHub Codespaces stack, then discusses prompt-injection safeguards, agent versus direct RAG, and Langfuse-based observability and session telemetry."
  },
  {
    "slug": "can-llms-generate-enterprise-quality-code-prasenjit-sarkar-sonar",
    "title": "Can LLMs generate Enterprise Quality Code? — Prasenjit Sarkar, Sonar",
    "url": "https://ai.engineer/talks/can-llms-generate-enterprise-quality-code-prasenjit-sarkar-sonar",
    "videoId": "NuePCNMpWGc",
    "event": "AI Engineer Europe 2026",
    "durationMs": 912000,
    "speakers": [
      {
        "slug": "prasenjit-sarkar",
        "name": "Prasenjit Sarkar",
        "organization": "Sonar"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Sonar’s Prasenjit Sarkar argues that coding-benchmark pass rates do not establish whether AI-generated software is secure, maintainable, or enterprise-ready. He describes SonarQube analysis of 4,444 Java assignments, comparing model-specific code volume and security findings: Claude Sonnet 4.6 generated approximately 627,000 lines and 300 security issues per million lines, while GPT-5.4 variants generated approximately 1.2 million lines. He concludes with Sonar’s ACDC guide-verify-solve framework and SonarQube Agentic Analysis for assessing agent-generated code."
  },
  {
    "slug": "can-oncology-workflows-run-without-human-touch-anant-shankhdhar-risa-labs",
    "title": "Can Oncology Workflows Run Without Human Touch? - Anant Shankhdhar, Risa Labs",
    "url": "https://ai.engineer/talks/can-oncology-workflows-run-without-human-touch-anant-shankhdhar-risa-labs",
    "videoId": "_cVfz88_j7A",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1001000,
    "speakers": [
      {
        "slug": "anant-shankhdhar",
        "name": "Anant Shankhdhar",
        "organization": "RISA Labs"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Anant Shankhdhar explains how RISA Labs automates oncology prior-authorization workflows with collaborating eligibility, authorization, medical-necessity, and submission agents. The system routes insurance checks through APIs or robotic process automation, normalizes payer data, uses LLM-generated portal configurations, and corroborates authorization decisions against historical records and a SQL-backed payer-rule knowledge base. Because LLM extraction is nondeterministic and clinical cases can require judgment, uncertain or complex cases are escalated for human review."
  },
  {
    "slug": "case-study-deep-dive-telemedicine-support-agents-with-langgraph-mcp",
    "title": "Case Study + Deep Dive: Telemedicine Support Agents with LangGraph/MCP",
    "url": "https://ai.engineer/talks/case-study-deep-dive-telemedicine-support-agents-with-langgraph-mcp",
    "videoId": "sn79oS4MZFI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 6973000,
    "speakers": [
      {
        "slug": "dan-mason",
        "name": "Dan Mason",
        "organization": "Stride"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Stride AI practice lead Dan Mason presents an interactive workshop on an agentic telemedicine system developed with Avila Science to support patients through multi-day, at-home treatment while preserving clinician relationships and human escalation. He explains configurable treatment blueprints, persistent patient state, messaging, adaptive schedules, virtual operations associates, LangGraph orchestration, MCP-connected tools within a VPC, and evaluation and response-validation techniques."
  },
  {
    "slug": "challenges-in-high-performance-robotics-systems",
    "title": "Challenges in High Performance Robotics Systems",
    "url": "https://ai.engineer/talks/challenges-in-high-performance-robotics-systems",
    "videoId": "bCGbuyv8PMk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 763000,
    "speakers": [
      {
        "slug": "rishabh-garg",
        "name": "Rishabh Garg",
        "organization": "Tesla Optimus"
      }
    ],
    "topics": [
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Rishabh Garg examines robotics failures that resemble faulty control policies but originate in software, communication, and timing. Using a toy robot architecture, he shows how CAN bus bandwidth limits create control-loop delays, how multithreaded TX/RX pipelines can introduce desynchronization and message jitter, and how external transceivers plus candump reveal those failures. He closes with synchronization primitives, timing padding, careful logging, and avoiding priority inversion."
  },
  {
    "slug": "challenges-to-scaling-agents-for-generative-ai-products",
    "title": "Challenges to Scaling Agents for Generative AI Products",
    "url": "https://ai.engineer/talks/challenges-to-scaling-agents-for-generative-ai-products",
    "videoId": "b2GqTDWtg6s",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1178000,
    "speakers": [
      {
        "slug": "anju-kambadur",
        "name": "Anju Kambadur",
        "organization": "Bloomberg"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Bloomberg AI Engineering leader Anju Kambadur traces the company's financial-domain language-model work and explains how it builds agentic products for research analysts. He describes cross-functional AI organization, structured and unstructured financial data, transparency and factuality requirements, MLOps remediation and circuit breakers, and why stochastic errors compound across composed LLM agents. He recommends downstream safety checks and highlights the evolution from sparse to dense and hybrid indexing."
  },
  {
    "slug": "chat-and-citations-won-t-save-your-vertical-ai",
    "title": "Chat and citations won't save your vertical AI",
    "url": "https://ai.engineer/talks/chat-and-citations-won-t-save-your-vertical-ai",
    "videoId": "RGiXcVxSD3s",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 912000,
    "speakers": [
      {
        "slug": "atul-ramachandran",
        "name": "Atul Ramachandran",
        "organization": "Filed"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Filed co-founder and CTO Atul Ramachandran argues that chat interfaces and citations alone do not make vertical AI useful when customers still shoulder verification and execution. Drawing on tax-preparation workflows, he proposes designing products around agentic delegation: users supervise AI workers, teach reusable skills, monitor long-running tasks through task lists and traces, and retain control when intervention is necessary."
  },
  {
    "slug": "chatgpt-is-poorly-designed-so-i-fixed-it",
    "title": "ChatGPT is poorly designed. So I fixed it",
    "url": "https://ai.engineer/talks/chatgpt-is-poorly-designed-so-i-fixed-it",
    "videoId": "y6L5RkEqQ8g",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 321000,
    "speakers": [
      {
        "slug": "ben-holmes",
        "name": "Ben Holmes",
        "organization": "Warp"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Ben Holmes critiques ChatGPT's separation of voice and text and confusing model selection, then demonstrates FixGPT, an open-source prototype combining realtime audio, persistent written responses, and heuristic routing to stronger reasoning models through tool calls. He relates the redesign to product fragmentation, interface patterns from messaging apps, and model-selection techniques used in Warp."
  },
  {
    "slug": "ci-cd-is-dead-agents-need-continuous-compute-and-computers-hugo-santos-and-madison-faulkner",
    "title": "CI/CD Is Dead, Agents Need Continuous Compute and Computers — Hugo Santos and Madison Faulkner",
    "url": "https://ai.engineer/talks/ci-cd-is-dead-agents-need-continuous-compute-and-computers-hugo-santos-and-madison-faulkner",
    "videoId": "VktrqzQgytY",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1117000,
    "speakers": [
      {
        "slug": "hugo-santos",
        "name": "Hugo Santos",
        "organization": "Namespace"
      },
      {
        "slug": "madison-faulkner",
        "name": "Madison Faulkner",
        "organization": "New Enterprise Associates (NEA)"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "NEA investor Madison Faulkner and Namespace CEO Hugo Santos argue that CI/CD pipelines and pull-request review were designed for human development rates and become bottlenecks when autonomous agents generate many simultaneous changes. They propose continuous compute, faster build-and-test infrastructure layered onto existing systems such as GitHub Actions, agent-scale orchestration, and intent-driven validation loops that preserve human acceptance of proposed changes."
  },
  {
    "slug": "ci-in-the-era-of-ai-from-unit-tests-to-stochastic-evals",
    "title": "CI in the Era of AI: From Unit Tests to Stochastic Evals",
    "url": "https://ai.engineer/talks/ci-in-the-era-of-ai-from-unit-tests-to-stochastic-evals",
    "videoId": "WXy8Yy9xGss",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 890000,
    "speakers": [
      {
        "slug": "nathan-sobo",
        "name": "Nathan Sobo",
        "organization": "Zed"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Zed co-founder Nathan Sobo explains how the Rust-native, GPU-rendered editor extends rigorous deterministic CI into stochastic evaluations for agentic coding. He describes replayable concurrency simulations, SWE-bench-like end-to-end agent evaluations, focused regression tests, Tree-sitter-assisted search, and defenses against streaming-diff errors, malformed XML, indentation mismatches, and model-specific escaping."
  },
  {
    "slug": "ciam-for-ai-authn-authz-for-agents-michael-grinich-ceo-of-workos",
    "title": "CIAM for AI: Authn/Authz for Agents — Michael Grinich, CEO of WorkOS",
    "url": "https://ai.engineer/talks/ciam-for-ai-authn-authz-for-agents-michael-grinich-ceo-of-workos",
    "videoId": "D4Dswf-__RM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1213000,
    "speakers": [
      {
        "slug": "michael-grinich",
        "name": "Michael Grinich",
        "organization": "WorkOS"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "WorkOS founder and CEO Michael Grinich explains why autonomous agents need dedicated authentication, authorization, scoped access, secure credential handling, and human-linked auditability. Using an accidental production-database deletion scenario, he surveys shadow identities, signed JWT delegation chains, capability tokens inspired by Macaroons, OAuth scopes, OpenID Connect-related standards, and WorkOS AuthKit. The solo presentation concludes with a question from an unidentified additional participant about agent-adoption timelines."
  },
  {
    "slug": "citation-needed-provenance-for-llm-built-knowledge-graphs",
    "title": "Citation Needed: Provenance for LLM-Built Knowledge Graphs",
    "url": "https://ai.engineer/talks/citation-needed-provenance-for-llm-built-knowledge-graphs",
    "videoId": "H7puB0RwJMM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1254000,
    "speakers": [
      {
        "slug": "daniel-chalef",
        "name": "Daniel Chalef",
        "organization": "Zep AI"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      }
    ],
    "summary": "Daniel Chalef explains why LLM-generated knowledge graphs lose reliable source lineage when facts are synthesized, entities merge, and newer information invalidates earlier claims. Using a healthcare example, he describes Graphiti's episode-linked graph provenance, source metadata tagging, temporal fact invalidation, compliance-oriented deletion, and debugging benefits. Audience questions address graph construction, episode ingestion, vector and full-text search, and breadth-first graph traversal."
  },
  {
    "slug": "claude-code-the-evolution-of-agentic-coding",
    "title": "Claude Code & the evolution of agentic coding",
    "url": "https://ai.engineer/talks/claude-code-the-evolution-of-agentic-coding",
    "videoId": "Lue8K2jqfKk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1092000,
    "speakers": [
      {
        "slug": "boris-cherny",
        "name": "Boris Cherny",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Claude Code creator Boris Cherny explains Anthropic's deliberately unopinionated approach to agentic coding as model capabilities and developer interfaces evolve rapidly. He traces programming's progression from punch cards and early editors to modern languages, then describes Claude Code installation and practical workflows including Unix-style model composition, codebase question answering, Git-history inspection, and test-driven development. A brief moderator-led audience question addresses managing multiple concurrent coding agents."
  },
  {
    "slug": "claude-fable-claude-tag-and-anthropic-s-culture-cat-wu-thariq-shihipar-ft-simon-willison",
    "title": "Claude Fable, Claude Tag, and Anthropic's Culture — Cat Wu & Thariq Shihipar ft Simon Willison",
    "url": "https://ai.engineer/talks/claude-fable-claude-tag-and-anthropic-s-culture-cat-wu-thariq-shihipar-ft-simon-willison",
    "videoId": "uU5Gv2h8-9g",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 3090000,
    "speakers": [
      {
        "slug": "cat-wu",
        "name": "Cat Wu",
        "organization": "Anthropic"
      },
      {
        "slug": "thariq-shihipar",
        "name": "Thariq Shihipar",
        "organization": "Anthropic"
      },
      {
        "slug": "simon-willison",
        "name": "Simon Willison",
        "organization": "Datasette"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Simon Willison hosts Anthropic's Cat Wu and Thariq Shihipar for a fireside chat about Claude Code, Claude Fable, and Claude Tag. They describe coding agents' shift from closely supervised implementation toward delegation, product judgment, collaborative workflows, software rewrites backed by strong tests, and changing engineering roles. The conversation covers an internal claim that Claude Tag lands 65% of product-engineering pull requests, code-review oversight, system-prompt reductions, agent-tool design, creative applications, and audience questions about evaluations, observability, memory, and multiplayer agents."
  },
  {
    "slug": "claude-for-long-horizon-tasks",
    "title": "Claude for long-horizon tasks",
    "url": "https://ai.engineer/talks/claude-for-long-horizon-tasks",
    "videoId": "9QebvrrY3KY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1519000,
    "speakers": [
      {
        "slug": "lance-martin",
        "name": "Lance Martin",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Anthropic's Lance Martin explains how increasing Claude task horizons change agent architecture, contrasting the Messages API, Claude Agent SDK, and Claude Managed Agents. He discusses externalized session context, separate-context verification, OpenAI's Parameter Golf, in-band agent memory, SQL-based continual-learning evaluation, and out-of-band dreaming that consolidates and corrects memory. The talk concludes with organization-level asynchronous agents and audience questions."
  },
  {
    "slug": "claude-plays-minecraft-introducing-a-real-world-serverless-ai-to-a-virtual-world",
    "title": "Claude plays Minecraft: Introducing a real world serverless AI to a virtual world",
    "url": "https://ai.engineer/talks/claude-plays-minecraft-introducing-a-real-world-serverless-ai-to-a-virtual-world",
    "videoId": "1B9i7FBsRVQ",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1096000,
    "speakers": [
      {
        "slug": "derek-bingham",
        "name": "Derek Bingham",
        "organization": "AWS"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "AWS developer advocate Derek Bingham demonstrates Rocky, an autonomous Minecraft agent that translates in-game chat into tool-driven actions such as navigation and digging. He outlines an initial LangChain, Lambda, SageMaker and Cohere implementation before describing a more serverless architecture using Agents for Amazon Bedrock, Mineflayer and Amazon ECS, alongside Bedrock knowledge bases, RAG and guardrails."
  },
  {
    "slug": "claws-out-securing-and-building-with-openclaw",
    "title": "Claws Out: Securing and Building with OpenClaw",
    "url": "https://ai.engineer/talks/claws-out-securing-and-building-with-openclaw",
    "videoId": "xg1zNlzw7Jk",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1032000,
    "speakers": [
      {
        "slug": "nick-taylor",
        "name": "Nick Taylor",
        "organization": "Pomerium"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Pomerium developer advocate Nick Taylor explains how he contributed trusted-proxy authentication to OpenClaw to protect its gateway behind an identity-aware proxy and avoid repeatedly supplying WebSocket tokens. He covers trusted proxy IPs, authenticated-user headers, JWTs, required headers, and policy-based authorization, then demonstrates building an MCP application using conference speaker data from an OpenClaw workspace."
  },
  {
    "slug": "climbing-the-ladder-of-abstraction",
    "title": "Climbing the Ladder of Abstraction",
    "url": "https://ai.engineer/talks/climbing-the-ladder-of-abstraction",
    "videoId": "PAy_GHUAICw",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1007000,
    "speakers": [
      {
        "slug": "amelia-wattenberger",
        "name": "Amelia Wattenberger",
        "organization": "Adept"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Amelia Wattenberger, speaking as a designer at Adept, argues that AI products should transform how people work by helping them move fluidly between levels of abstraction rather than merely accelerating existing tasks. Using spreadsheets and Google Maps as precedents, she reframes augmentation as many smaller automations, demonstrates LLM-generated summaries and writing views, and imagines software-operating AI that helps users compare travel listings, extract relevant details, and initiate actions while remaining in control."
  },
  {
    "slug": "code-generation-and-maintenance-at-scale",
    "title": "Code Generation and Maintenance at Scale",
    "url": "https://ai.engineer/talks/code-generation-and-maintenance-at-scale",
    "videoId": "Ve-akpov78Q",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1134000,
    "speakers": [
      {
        "slug": "morgante-pell",
        "name": "Morgante Pell",
        "organization": "Grit"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Grit founder Morgante Pell argues that effective AI coding agents should help expert engineers maintain and migrate existing enterprise codebases rather than merely generate new applications. He describes coordinating large-scale pull-request workflows, illustrates organizational migration challenges with OpenTelemetry, and explains how language-model agents benefit from sub-agent delegation, TypeScript server feedback, incremental in-memory analysis, and more precise code-editing interfaces."
  },
  {
    "slug": "code-mode-let-the-code-do-the-talking",
    "title": "Code Mode: Let the Code do the Talking",
    "url": "https://ai.engineer/talks/code-mode-let-the-code-do-the-talking",
    "videoId": "8txf05vVVl4",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1180000,
    "speakers": [
      {
        "slug": "sunil-pai",
        "name": "Sunil Pai",
        "organization": "Cloudflare"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Cloudflare’s Sunil Pai presents Code Mode, an agent architecture in which models generate executable JavaScript against typed APIs instead of repeatedly exchanging JSON tool calls. He explains how OpenAPI-driven search and execution can expose large API surfaces without overwhelming model context, demonstrates discovering and listing Cloudflare Workers, and discusses composed stateful operations, sandbox restrictions, observability, generative interfaces, and designing developer experiences for agents."
  },
  {
    "slug": "code-world-model-building-world-models-for-computation",
    "title": "Code World Model: Building World Models for Computation",
    "url": "https://ai.engineer/talks/code-world-model-building-world-models-for-computation",
    "videoId": "sYgE4ppDFOQ",
    "event": "AI Engineer Code 2025",
    "durationMs": 1001000,
    "speakers": [
      {
        "slug": "jacob-kahn",
        "name": "Jacob Kahn",
        "organization": "FAIR, Meta"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "Meta FAIR researcher Jacob Kahn presents Code World Model, a 32-billion-parameter dense research model that learns computational behavior from program execution traces and repository-level GitHub data. He explains observation-action world modeling, multistage training with supervised fine-tuning and reinforcement learning, and Bash-centered software-engineering agents that reason about code, modify repositories, and learn from environmental feedback."
  },
  {
    "slug": "codex-and-subagents",
    "title": "Codex and Subagents",
    "url": "https://ai.engineer/talks/codex-and-subagents",
    "videoId": "MhHEGMFCEB0",
    "event": "AI Engineer Europe 2026",
    "durationMs": 3719000,
    "speakers": [
      {
        "slug": "vaibhav-srivastav",
        "name": "Vaibhav Srivastav",
        "organization": "OpenAI"
      },
      {
        "slug": "katia-gil-guzman",
        "name": "Katia Gil Guzman",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "OpenAI presenters Vaibhav Srivastav and Katia Gil Guzman demonstrate Codex as an agentic software engineering system, covering its GPT model options, recurring automations, Slack integrations, and plugins combining skills, apps, and MCP servers. They show specialized custom subagents for reviewing code and agent personas, describe event-driven GitHub hooks, and answer audience questions about session history and cloud support."
  },
  {
    "slug": "codex-behind-the-harness",
    "title": "Codex, Behind the Harness",
    "url": "https://ai.engineer/talks/codex-behind-the-harness",
    "videoId": "shRR1e2HXMk",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1255000,
    "speakers": [
      {
        "slug": "dominik-kundel",
        "name": "Dominik Kundel",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "OpenAI’s Dominik Kundel dissects the open-source, Rust-based Codex agent harness, explaining how the app-server connects user interfaces to the harness while the Responses API connects it to model inference and tools. He describes interoperable providers, context and skills budgeting, MCP integrations, file system interactions, and a TypeScript nano Codex demonstration, with examples embedding Codex in other interfaces."
  },
  {
    "slug": "coding-evals-from-code-snippets-to-codebases-naman-jain-cursor",
    "title": "Coding Evals: From Code Snippets to Codebases – Naman Jain, Cursor",
    "url": "https://ai.engineer/talks/coding-evals-from-code-snippets-to-codebases-naman-jain-cursor",
    "videoId": "tHN44yJoeS8",
    "event": "AI Engineer Code 2025",
    "durationMs": 1088000,
    "speakers": [
      {
        "slug": "naman-jain",
        "name": "Naman Jain",
        "organization": "Cursor"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Cursor engineer Naman Jain traces coding-model evaluation from single-line completions and competitive-programming tasks to repository-scale software optimization. He explains how LiveCodeBench uses continuously refreshed problems to expose contamination and changing difficulty, then discusses SWE-bench-style agent workloads, optimizing llama.cpp with Qwen, infrastructure and test manipulation, complex codebases such as Google’s Zopfli, and intermediate grading signals for longer-horizon tasks."
  },
  {
    "slug": "cognitive-exhaust-fumes-or-read-only-ai-is-underrated-simon-podhajsky-head-of-ai-waypoint",
    "title": "Cognitive Exhaust Fumes, or: Read-Only AI Is Underrated — Šimon Podhajský, Head of AI, Waypoint",
    "url": "https://ai.engineer/talks/cognitive-exhaust-fumes-or-read-only-ai-is-underrated-simon-podhajsky-head-of-ai-waypoint",
    "videoId": "u0TOSBbAw7c",
    "event": "AI Engineer Europe 2026",
    "durationMs": 691000,
    "speakers": [
      {
        "slug": "simon-podhajsky",
        "name": "Šimon Podhajský",
        "organization": "Waypoint AI"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Šimon Podhajský presents a personal AI observer that has read-only access to six data sources and writes analyses only to an isolated review workspace. Using Claude, Python, the Anthropic API, browser-history data, and Clay MCP, he demonstrates weekly reflections and recommendations for whom to discuss recent reading with. He argues that combining digital traces reveals patterns individual applications cannot, while prohibiting source writes reduces consequential errors, prevents contamination of behavioral evidence, and preserves human agency."
  },
  {
    "slug": "cognitive-shield-real-time-real-smart-rachna-srivastava",
    "title": "Cognitive Shield Real Time Real Smart - Rachna Srivastava",
    "url": "https://ai.engineer/talks/cognitive-shield-real-time-real-smart-rachna-srivastava",
    "videoId": "5_QWh4LGoxg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2698000,
    "speakers": [
      {
        "slug": "rachna-srivastava",
        "name": "Rachna Srivastava"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Rachna Srivastava presents Cognitive Shield, a three-layer financial-fraud defense platform addressing voice-cloning scams, deepfake identity verification, romance-investment fraud, and synthetic cryptocurrency promotions. The demonstration describes Neo4j-backed fraud graphs, natural-language Cypher query generation, GraphRAG, and a multimodal investigation assistant, while emphasizing explainable decisions and privacy by design."
  },
  {
    "slug": "cohere-for-vps-of-ai",
    "title": "Cohere for VPs of AI",
    "url": "https://ai.engineer/talks/cohere-for-vps-of-ai",
    "videoId": "u3NofYYstaY",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 971000,
    "speakers": [
      {
        "slug": "vivek-muppalla",
        "name": "Vivek Muppalla",
        "organization": "Cohere"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Cohere engineering director Vivek Muppalla briefs AI leaders on the company's enterprise AI offering, contrasting Command R and Command R+ and describing embeddings, reranking, retrieval-augmented generation with built-in citations, multilingual evaluation, and agentic tool use. He also discusses enterprise data privacy, intellectual-property indemnification, and deployment across clouds, private infrastructure, and on-premises environments before audience questions on text classification and model evolution."
  },
  {
    "slug": "collaborating-with-agents-in-your-software-development-workflow-jon-peck-christopher-harrison-gi",
    "title": "Collaborating with Agents in Your Software Development Workflow - Jon Peck & Christopher Harrison, GitHub",
    "url": "https://ai.engineer/talks/collaborating-with-agents-in-your-software-development-workflow-jon-peck-christopher-harrison-gi",
    "videoId": "G1hhmz6mXT0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 3846000,
    "speakers": [
      {
        "slug": "jon-peck",
        "name": "Jon Peck",
        "organization": "GitHub"
      },
      {
        "slug": "christopher-harrison",
        "name": "Christopher Harrison",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "GitHub enterprise advocates Christopher Harrison and Jon Peck lead a hands-on workshop on incorporating GitHub Copilot agents into software development. They cover participant onboarding and repository setup, framing Copilot as an AI pair programmer, supplying explicit task context, extending coding agents and Copilot Chat with MCP servers, maintaining repository instruction files, and integrating Azure MCP Server. Audience questions address agent behavior, enterprise integrations, and MCP-server configuration."
  },
  {
    "slug": "combine-skills-and-mcp-to-close-the-context-gap",
    "title": "Combine Skills and MCP to Close the Context Gap",
    "url": "https://ai.engineer/talks/combine-skills-and-mcp-to-close-the-context-gap",
    "videoId": "JT3OzDKrucU",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1107000,
    "speakers": [
      {
        "slug": "pedro-rodrigues",
        "name": "Pedro Rodrigues",
        "organization": "Supabase"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Supabase AI tooling engineer Pedro Rodrigues explains why MCP tools need product-specific agent skills to operate safely and reliably. A PostgreSQL example shows an agent creating a view that can bypass row-level security unless security_invoker = true is set. He describes Supabase’s agent skill, SSH-accessible documentation, opinionated development workflows, and evaluations across Claude and GPT models in which MCP plus skills outperformed baseline and MCP-only conditions. Audience questions address vector databases and distributing skills within organizations."
  },
  {
    "slug": "comfyui-workshop-with-comfyanonymous-and-jedrick-kosinski",
    "title": "ComfyUI Workshop with ComfyAnonymous and Jedrick Kosinski",
    "url": "https://ai.engineer/talks/comfyui-workshop-with-comfyanonymous-and-jedrick-kosinski",
    "videoId": "_FKeSzM9fPc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 3085000,
    "speakers": [
      {
        "slug": "jedrick-kosinski",
        "name": "Jedrick Kosinski",
        "organization": "Comfy Org"
      },
      {
        "slug": "comfyanonymous",
        "name": "ComfyAnonymous",
        "organization": "Comfy Org"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "Jedrick Kosinski and ComfyUI creator ComfyAnonymous introduce Comfy Org's open-source, node-based platform for image, video, audio, 3D, and other generative AI workflows. They explain local and API-based model execution, reproducible workflows embedded in generated media, diffusion-model prompting and classifier-free guidance, and FLUX.1 Kontext. An extended audience Q&A explores image generation, workflow APIs, product-roadmap uncertainty, custom nodes, and potential cloud inference."
  },
  {
    "slug": "compilers-in-the-age-of-llms",
    "title": "Compilers in the Age of LLMs",
    "url": "https://ai.engineer/talks/compilers-in-the-age-of-llms",
    "videoId": "q2nHsJVy4FE",
    "event": "AI Engineer Code 2025",
    "durationMs": 1056000,
    "speakers": [
      {
        "slug": "yusuf-olokoba",
        "name": "Yusuf Olokoba",
        "organization": "Muna"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Yusuf Olokoba of Muna presents a compiler that turns ordinary Python AI workloads into portable native binaries while exposing open-source models through an OpenAI-compatible client. He argues for hybrid cloud, edge, and on-device inference, demonstrates an EmbeddingGemma-based embedding workflow, explains why PyTorch tracing was insufficient for arbitrary Python dependencies, and describes using LLMs to generate C++ and Rust implementations of compiler operations. Google's official documentation contradicts the talk's stated EmbeddingGemma parameter count: the model has 308 million parameters, not 270 million."
  },
  {
    "slug": "comprehend-first-code-later-the-ai-skill-i-rely-on-daily",
    "title": "Comprehend First, Code Later: The AI Skill I Rely On Daily",
    "url": "https://ai.engineer/talks/comprehend-first-code-later-the-ai-skill-i-rely-on-daily",
    "videoId": "li0SaBt9RDM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1025000,
    "speakers": [
      {
        "slug": "priscila-andre-de-oliveira",
        "name": "Priscila Andre de Oliveira",
        "organization": "Sentry"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Sentry senior software engineer Priscila Andre de Oliveira argues that AI's greatest day-to-day engineering benefit is understanding complex code rather than generating it. She describes Sentry's internal AI tools and quality practices, demonstrates a Catch Me Up repository-comprehension skill using Claude Opus, and recommends understanding an agent's findings before planning or implementation. In her tracked usage, 67% of prompts concerned comprehension and only 2% concerned generation."
  },
  {
    "slug": "compression-at-the-edge",
    "title": "Compression at the Edge",
    "url": "https://ai.engineer/talks/compression-at-the-edge",
    "videoId": "J4_jCrTxMkk",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2761000,
    "speakers": [
      {
        "slug": "chris-alexiuk",
        "name": "Chris Alexiuk",
        "organization": "NVIDIA"
      },
      {
        "slug": "daniel-han",
        "name": "Daniel Han",
        "organization": "Unsloth"
      },
      {
        "slug": "asma-beevi",
        "name": "Asma Beevi",
        "organization": "NVIDIA"
      },
      {
        "slug": "merve-noyan",
        "name": "Merve Noyan",
        "organization": "Hugging Face"
      },
      {
        "slug": "parth-sareen",
        "name": "Parth Sareen",
        "organization": "Ollama"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Moderator Chris Alexiuk leads a panel with representatives from NVIDIA, Unsloth, Hugging Face, and Ollama on compressing open models for local and edge deployment. The discussion covers FP32-to-FP4 and mixed-precision quantization, model-size versus quality tradeoffs, DeepSeek-R1, layer-sensitive compression, linear-attention limitations, long-context evaluation, and benchmarking compressed models."
  },
  {
    "slug": "compute-system-design-for-next-generation-frontier-models",
    "title": "Compute & System Design for Next Generation Frontier Models",
    "url": "https://ai.engineer/talks/compute-system-design-for-next-generation-frontier-models",
    "videoId": "gFyBdBm0AGo",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1109000,
    "speakers": [
      {
        "slug": "dylan-patel",
        "name": "Dylan Patel",
        "organization": "SemiAnalysis"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "SemiAnalysis founder Dylan Patel examines the infrastructure required to serve increasingly capable frontier models, contrasting compute-intensive prompt prefill with bandwidth-intensive token decoding. He discusses Llama 405B deployment, vLLM, TensorRT-LLM, continuous batching, disaggregated prefill, accelerator contention, service reliability, context caching, enterprise document-processing costs, and the scale of frontier-model data centers."
  },
  {
    "slug": "computer-use-2-0-agents-just-got-multi-cursor",
    "title": "Computer-Use 2.0: Agents Just Got Multi-Cursor",
    "url": "https://ai.engineer/talks/computer-use-2-0-agents-just-got-multi-cursor",
    "videoId": "ZSQb5fzRFPw",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1001000,
    "speakers": [
      {
        "slug": "francesco-bonacci",
        "name": "Francesco Bonacci",
        "organization": "Cua"
      },
      {
        "slug": "dillon-dupont",
        "name": "Dillon DuPont",
        "organization": "Cua"
      },
      {
        "slug": "robert-wendt",
        "name": "Robert Wendt",
        "organization": "Cua"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Francesco Bonacci, Dillon DuPont, and Robert Wendt present Cua's stack for computer-use agents: Cua Driver enables background, cross-platform desktop interaction through accessibility trees and alternative click paths; Cua-Bench evaluates agents across more than 130 tasks, 42 environments, and five platforms; a Snorkel AI collaboration exposes failures on blank KiCad schematics; and pooled infrastructure targets idle-GPU costs during reinforcement-learning workloads."
  },
  {
    "slug": "computer-use-at-the-edge-of-the-statistical-precipice",
    "title": "Computer Use at the Edge of the Statistical Precipice",
    "url": "https://ai.engineer/talks/computer-use-at-the-edge-of-the-statistical-precipice",
    "videoId": "CTLa_p6iOiY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1048000,
    "speakers": [
      {
        "slug": "pierluca-d-oro",
        "name": "Pierluca D'Oro",
        "organization": "Programma Labs"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Pierluca D'Oro demonstrates that small blind replay scripts can match or outperform frontier agents on deterministic computer-use benchmarks and that their expected success rate corresponds to pass@k. He proposes PRISM principles for trustworthy environments and introduces DigiWorld, an Android benchmark spanning 15 apps, 387 verified scenarios, and 3.2 million verified configurations generated through a compiler-like pipeline. He also shows that rollout-only nominal 95% confidence intervals can achieve only approximately 17–20% coverage, motivating more rigorous evaluation and uncertainty estimation."
  },
  {
    "slug": "computer-use-models-will-agentify-the-web-not-apis",
    "title": "Computer-use models will agentify the web, not APIs",
    "url": "https://ai.engineer/talks/computer-use-models-will-agentify-the-web-not-apis",
    "videoId": "Ki980nV0__0",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1260000,
    "speakers": [
      {
        "slug": "dhruv-batra",
        "name": "Dhruv Batra",
        "organization": "Yutori"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Yutori co-founder and chief scientist Dhruv Batra argues that most websites will not expose agent-ready APIs, making computer-use models essential for navigating the web’s long tail. Restaurant menus embedded in images, scanned procurement documents, and Freedom of Information Act workflows illustrate why text extraction and structured interfaces are insufficient. He explains that browser state is often visible only after rendering, demonstrates screenshot-grounded agents that can also execute JavaScript, and discusses Online-Mind2Web evaluation alongside latency and cost per task."
  },
  {
    "slug": "connecting-the-dots-with-context-graphs",
    "title": "Connecting the Dots with Context Graphs",
    "url": "https://ai.engineer/talks/connecting-the-dots-with-context-graphs",
    "videoId": "eW_vxrjvERk",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1059000,
    "speakers": [
      {
        "slug": "stephen-chin",
        "name": "Stephen Chin",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      }
    ],
    "summary": "Neo4j developer-relations leader Stephen Chin explains how context graphs combine knowledge graphs, vector retrieval, and durable agent memory to ground AI decisions in relationships and prior reasoning. A healthcare example contrasts generic LLM and vector-RAG responses with graph-grounded context; subsequent sections cover Cypher, multi-hop traversal, FastRP embeddings, Louvain community grouping, and the open-source Neo4j Agent Memory project. Podcast and financial-services demonstrations illustrate searchable memory, decision history, and explainable workflows."
  },
  {
    "slug": "conquering-agent-chaos",
    "title": "Conquering Agent Chaos",
    "url": "https://ai.engineer/talks/conquering-agent-chaos",
    "videoId": "yASxPZ-tZe0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 880000,
    "speakers": [
      {
        "slug": "rick-blalock",
        "name": "Rick Blalock",
        "organization": "Agentuity"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Agentuity founder Rick Blalock explains why deploying long-running AI agents on conventional serverless infrastructure can fail through execution timeouts, networking limitations, and operational complexity. He demonstrates Agentuity's CLI, Bun/Python/Node.js runtime options, framework-agnostic templates including Vercel AI SDK and Groq, agent routing and API-key controls, GitHub-triggered deployment, and built-in observability and AI-gateway capabilities."
  },
  {
    "slug": "containing-agent-chaos",
    "title": "Containing Agent Chaos",
    "url": "https://ai.engineer/talks/containing-agent-chaos",
    "videoId": "bUBF5V6oDKw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1428000,
    "speakers": [
      {
        "slug": "solomon-hykes",
        "name": "Solomon Hykes",
        "organization": "DAgger"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Solomon Hykes presents Dagger's approach to controlling autonomous coding-agent chaos through isolated container environments, asynchronous execution, explicit guardrails, and efficient human intervention. He demonstrates Container Use connecting Claude Code through MCP, supporting model-independent workflows, parallel experiments, Git-based inspection, and execution across remote infrastructure or CI."
  },
  {
    "slug": "content-is-code",
    "title": "Content Is Code",
    "url": "https://ai.engineer/talks/content-is-code",
    "videoId": "yv6xovSsB1U",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 653000,
    "speakers": [
      {
        "slug": "matt-palmer",
        "name": "Matt Palmer",
        "organization": "Conductor"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Matt Palmer explains why AI makes code a practical source of truth for technical communication, including documentation, changelogs, product updates, websites, and video. Using a Conductor product walkthrough built with React and Remotion, he describes TypeScript-based content engineering and argues that clean pull requests, organizational structure, and conscientiousness matter more than indiscriminate AI generation."
  },
  {
    "slug": "context-engineering-connecting-the-dots-with-graphs-stephen-chin-neo4j",
    "title": "Context Engineering: Connecting the Dots with Graphs — Stephen Chin, Neo4j",
    "url": "https://ai.engineer/talks/context-engineering-connecting-the-dots-with-graphs-stephen-chin-neo4j",
    "videoId": "LLuKshphGOE",
    "event": "AI Engineer Code 2025",
    "durationMs": 1610000,
    "speakers": [
      {
        "slug": "stephen-chin",
        "name": "Stephen Chin",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Neo4j developer-relations leader Stephen Chin explains how context engineering improves AI applications by combining enterprise retrieval, curated agent memory, knowledge graphs, and GraphRAG. He discusses context-window limitations, hallucination reduction, dynamic prompting with DSPy and BAML, graph relationships and embeddings, role-sensitive access to information, vulnerability-document ingestion, and time-aware context for LLMs."
  },
  {
    "slug": "context-engineering-in-2026-compaction-memory-cost",
    "title": "Context Engineering in 2026: Compaction, Memory & Cost",
    "url": "https://ai.engineer/talks/context-engineering-in-2026-compaction-memory-cost",
    "videoId": "WP3hjUXd918",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 3806000,
    "speakers": [
      {
        "slug": "louis-francois-bouchard",
        "name": "Louis-François Bouchard",
        "organization": "Towards AI"
      },
      {
        "slug": "omar-solano",
        "name": "Omar Solano",
        "organization": "Towards AI"
      },
      {
        "slug": "samridhi-vaid",
        "name": "Samridhi Vaid",
        "organization": "Towards AI"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Towards AI presenters Louis-François Bouchard, Omar Solano, and Samridhi Vaid demonstrate how they evaluate and optimize an open-source AI tutor’s conversational context. The workshop covers context rot, compaction and delta summarization, caching and full-history baselines, multi-turn recall, Hugging Face deployment, Opik observability, latency, time to first token, and the cost of evaluation runs."
  },
  {
    "slug": "context-graphs-for-explainable-decision-aware-ai-agents",
    "title": "Context Graphs for Explainable, Decision-Aware AI Agents",
    "url": "https://ai.engineer/talks/context-graphs-for-explainable-decision-aware-ai-agents",
    "videoId": "abvQEhvRI_c",
    "event": "AI Engineer Europe 2026",
    "durationMs": 999000,
    "speakers": [
      {
        "slug": "andreas-kollegger",
        "name": "Andreas Kollegger",
        "organization": "Neo4j"
      },
      {
        "slug": "zaid-zaim",
        "name": "Zaid Zaim",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Neo4j presenters explain how context graphs extend knowledge graphs with short-term, long-term, and reasoning memory so AI agents can make explainable decisions grounded in organizational context, policies, and rules. They outline agentic GraphRAG and Text2Cypher graph traversal, then examine an agent decision framework, including the risks of autonomous actions, healthcare examples, the cost of incorrect decisions, and the need for domain-specific implementation."
  },
  {
    "slug": "context-is-the-new-code",
    "title": "Context Is the New Code",
    "url": "https://ai.engineer/talks/context-is-the-new-code",
    "videoId": "bSG9wUYaHWU",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1634000,
    "speakers": [
      {
        "slug": "patrick-debois",
        "name": "Patrick Debois",
        "organization": "Tessl"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Patrick Debois of Tessl argues that the instructions, skills, documentation, and organizational knowledge guiding AI coding agents deserve the same engineering discipline as source code. He outlines a Context Development Lifecycle covering generation, evaluation, distribution, and observation; discusses AGENTS.md, CLAUDE.md, and MCP-connected tools; and explains how tests, linters, reusable skills, instrumentation, and feedback can improve agent context over time. The presentation concludes with audience questions."
  },
  {
    "slug": "context-platform-engineering-to-reduce-token-anxiety-val-bercovici-and-callan-fox-weka",
    "title": "Context Platform Engineering to Reduce Token Anxiety — Val Bercovici and Callan Fox, WEKA",
    "url": "https://ai.engineer/talks/context-platform-engineering-to-reduce-token-anxiety-val-bercovici-and-callan-fox-weka",
    "videoId": "NTBX-wxUhHs",
    "event": "AI Engineer Code 2025",
    "durationMs": 1432000,
    "speakers": [
      {
        "slug": "val-bercovici",
        "name": "Val Bercovici",
        "organization": "WEKA"
      },
      {
        "slug": "callan-fox",
        "name": "Callan Fox",
        "organization": "WEKA"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "WEKA presenters Val Bercovici and Callan Fox introduce context-platform engineering for agent swarms, including an open-source load-generation toolkit, KV-cache optimization, and translation of agent SLAs into infrastructure SLOs. Fox examines context-window saturation, repeated prefills and token costs, then discusses GPU pressure and memory-tier design, including pooled DRAM and persistent KV-cache infrastructure."
  },
  {
    "slug": "continual-learning-for-ai-agents-from-failures-to-durable-improvements-soheil-feizi-relai",
    "title": "Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI",
    "url": "https://ai.engineer/talks/continual-learning-for-ai-agents-from-failures-to-durable-improvements-soheil-feizi-relai",
    "videoId": "2IxD9OB3XuQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1355000,
    "speakers": [
      {
        "slug": "soheil-feizi",
        "name": "Soheil Feizi",
        "organization": "RELAI; University of Maryland, College Park"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Soheil Feizi, RELAI founder and chief scientific officer and a University of Maryland computer science associate professor, presents verifiable continual learning for AI agents. He explains how production failures and feedback can drive improvements across model, harness, and memory layers, comparing supervised fine-tuning, reinforcement-learning methods, LoRA, and GEPA-style prompt optimization. His proposed learning loop turns failures into replayable environments, validates changes against regression tests, and produces reviewable updates guided by replayability, holisticness, lifelongness, and efficiency."
  },
  {
    "slug": "convex-launch",
    "title": "Convex Launch",
    "url": "https://ai.engineer/talks/convex-launch",
    "videoId": "JbyictbPFV0",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 336000,
    "speakers": [
      {
        "slug": "jamie-turner",
        "name": "Jamie Turner",
        "organization": "Convex"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Convex co-founder Jamie Turner explains how a backend platform designed around React-like reactive data flow, subscribable queries, mutations, and synchronized application state became well suited to multistep generative AI applications. He illustrates workflows involving speech recognition, summaries, embeddings, and generated project plans, then describes built-in vector indexes and reusable Convex Components for sophisticated stateful backend workflows."
  },
  {
    "slug": "cooking-with-agents-in-vs-code",
    "title": "Cooking with Agents in VS Code",
    "url": "https://ai.engineer/talks/cooking-with-agents-in-vs-code",
    "videoId": "dyHpnnlkTc8",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1024000,
    "speakers": [
      {
        "slug": "liam-hampton",
        "name": "Liam Hampton",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Microsoft cloud advocate Liam Hampton demonstrates choosing among three VS Code coding-agent execution modes: interactive local agents for hands-on work, background agents using Copilot CLI and isolated Git worktrees for partially supervised tasks, and cloud agents for autonomous documentation and repository improvements. Using a Python CRUD application and a frontend issue, he emphasizes concurrent agent workflows, appropriate human oversight, token costs, and patterns transferable beyond GitHub Copilot."
  },
  {
    "slug": "cooking-with-fire-without-burning-down-the-kitchen",
    "title": "Cooking with fire without burning down the kitchen",
    "url": "https://ai.engineer/talks/cooking-with-fire-without-burning-down-the-kitchen",
    "videoId": "AVjrkXGnF2M",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1235000,
    "speakers": [
      {
        "slug": "dominik-kundel",
        "name": "Dominik Kundel",
        "organization": "Twilio"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Twilio product and design leader Dominik Kundel explains how a small Emerging Tech and Innovation team explores AI-driven customer engagement without disrupting existing customers. He contrasts sustaining product improvements with disruptive AI agents, highlights enterprise quality and cost limitations, and recommends early customer exposure, internal hackathons, rapid iteration, empowered developers, and continuous evaluation of new models."
  },
  {
    "slug": "copilots-everywhere",
    "title": "Copilots Everywhere",
    "url": "https://ai.engineer/talks/copilots-everywhere",
    "videoId": "nZEaxaKmTG0",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1102000,
    "speakers": [
      {
        "slug": "thomas-dohmke",
        "name": "Thomas Dohmke",
        "organization": "GitHub"
      },
      {
        "slug": "eugene-yan",
        "name": "Eugene Yan",
        "organization": "Amazon"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "In this closing-keynote conversation, Thomas Dohmke recounts how GitHub developed Copilot from early OpenAI Codex experiments during lockdown, validated it through internal developer feedback, and tracked generated-code adoption. He discusses extending Copilot throughout GitHub, adapting product strategy to rapid AI advances, and designing human-centered agents that augment developers, preserve institutional knowledge, and make software development more enjoyable."
  },
  {
    "slug": "covalent-launch-the-gpu-cheatcode-fine-tune-20-llama-models-in-5-minutes",
    "title": "Covalent Launch: The GPU Cheatcode: Fine-tune 20 Llama Models in 5 Minutes",
    "url": "https://ai.engineer/talks/covalent-launch-the-gpu-cheatcode-fine-tune-20-llama-models-in-5-minutes",
    "videoId": "zHYQZFy0UVk",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 386000,
    "speakers": [
      {
        "slug": "santosh-radha",
        "name": "Santosh Radha",
        "organization": "Agnostiq (Covalent)"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Santosh Radha introduces Agnostiq’s Covalent platform for orchestrating model fine-tuning, evaluation, and inference directly from Python without managing Docker or Kubernetes. He demonstrates dispatching notebook functions to remote GPUs, combining hosted or customer-provided compute with usage-based billing, and deploying inference endpoints with configurable autoscaling. The walkthrough assigns GPU and CPU resources to different workflow stages and discusses example evaluation costs."
  },
  {
    "slug": "crabrag-why-automated-assistants-need-graph-memory-not-more-tokens",
    "title": "CrabRAG: Why Automated Assistants Need Graph Memory, Not More Tokens",
    "url": "https://ai.engineer/talks/crabrag-why-automated-assistants-need-graph-memory-not-more-tokens",
    "videoId": "Q0VkgCyNVUg",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1242000,
    "speakers": [
      {
        "slug": "stephen-chin",
        "name": "Stephen Chin",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Neo4j developer-relations leader Stephen Chin argues that autonomous assistants need persistent graph-based memory because markdown files, standalone vector retrieval, and MCP-backed memory tools struggle with scalable, multi-hop context. He demonstrates a home-network digital twin in which vector-seeded Neo4j graph traversal and Cypher queries identify outdated software and exposed management interfaces, producing actionable, explainable results that similarity search misses."
  },
  {
    "slug": "creating-agents-that-co-create",
    "title": "Creating Agents That Co-Create",
    "url": "https://ai.engineer/talks/creating-agents-that-co-create",
    "videoId": "1XvN5EBDnDw",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1462000,
    "speakers": [
      {
        "slug": "karina-nguyen",
        "name": "Karina Nguyen",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "OpenAI researcher Karina Nguyen explains how large-scale pre-training, post-training with human and AI feedback, and reinforcement learning over chains of thought are transforming AI products into collaborative agents. Using GitHub Copilot, OpenAI o1, Canvas, and ChatGPT Tasks, she discusses coding assistance, long-horizon tool use, synthetic user data, personalization, familiar product interfaces, and a future of human-AI creative co-direction."
  },
  {
    "slug": "creating-and-scaling-your-own-custom-copilots-with-azure-ai-studio",
    "title": "Creating and scaling your own custom copilots with Azure AI Studio",
    "url": "https://ai.engineer/talks/creating-and-scaling-your-own-custom-copilots-with-azure-ai-studio",
    "videoId": "NjNEdjDvKu8",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1461000,
    "speakers": [
      {
        "slug": "hanchi-wang",
        "name": "Hanchi Wang",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Microsoft engineering lead Hanchi Wang demonstrates how Azure AI Studio and the open-source Prompt flow toolkit support building, tracing, evaluating, and monitoring custom enterprise copilots. A tool-using Assistant API example combines natural-language-to-SQL data retrieval with a code interpreter, while demonstrations cover Python trace instrumentation, local CLI testing, and custom evaluation metrics including execution time, error rate, and SQL similarity against ground truth."
  },
  {
    "slug": "critical-ai-inference-your-cio-can-trust",
    "title": "Critical AI Inference Your CIO Can Trust",
    "url": "https://ai.engineer/talks/critical-ai-inference-your-cio-can-trust",
    "videoId": "6Tpm4m1YxHk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1144000,
    "speakers": [
      {
        "slug": "sahil-yadav",
        "name": "Sahil Yadav",
        "organization": "Telemetrak"
      },
      {
        "slug": "hariharan-ganesan",
        "name": "Hariharan Ganesan",
        "organization": "Telemetrak"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Sahil Yadav and Hariharan Ganesan examine why enterprise AI inference must be trustworthy in telecom, industrial IoT, and other safety-critical environments. They outline three operational foundations—plain-language explainability, adaptive guardrails with human escalation, and digitally signed traceability—then describe XTOps trust dashboards, the MTTRE incident metric, and a Guardhat industrial-safety example involving GPS drift. The discussion closes by acknowledging the difficulty of quantifying the financial value of AI trust."
  },
  {
    "slug": "customized-production-ready-inference-with-open-source-models-dmytro-dima-dzhulgakov",
    "title": "Customized, production ready inference with open source models: Dmytro (Dima) Dzhulgakov",
    "url": "https://ai.engineer/talks/customized-production-ready-inference-with-open-source-models-dmytro-dima-dzhulgakov",
    "videoId": "ePMvfa8vgL8",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1135000,
    "speakers": [
      {
        "slug": "dmytro-dima-dzhulgakov",
        "name": "Dmytro (Dima) Dzhulgakov",
        "organization": "Fireworks AI"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Fireworks AI co-founder Dmytro Dzhulgakov substitutes for scheduled speaker Lin Qiao and explains how customizable open models can reduce inference latency and cost while improving domain-specific performance. He discusses fine-tuned Llama and Gemma models, agentic workloads and function calling, operational challenges around GPUs and production deployment, custom CUDA kernels, long-prompt RAG and caching, multimodal Stable Diffusion inference, and structured JSON generation."
  },
  {
    "slug": "dark-factory-openclaw-ships-faster-than-you-can-read-the-diff",
    "title": "Dark Factory: OpenClaw Ships Faster Than You Can Read the Diff",
    "url": "https://ai.engineer/talks/dark-factory-openclaw-ships-faster-than-you-can-read-the-diff",
    "videoId": "pmoDeA3RBZY",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1004000,
    "speakers": [
      {
        "slug": "vincent-koc",
        "name": "Vincent Koc",
        "organization": "Comet ML / OpenClaw"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "OpenClaw maintainer Vincent Koc describes a dark-factory approach to software engineering in which developers supervise many parallel coding agents instead of manually producing every change. Drawing on rapid OpenClaw development, collaboration around NVIDIA NemoClaw, and disruptive codebase refactoring, he argues that engineering judgment, agent oversight, modular integrations, and token efficiency matter more than maximizing raw commits or token consumption."
  },
  {
    "slug": "data-and-environment-curation-for-post-training-llms",
    "title": "Data and Environment Curation for Post-training LLMs",
    "url": "https://ai.engineer/talks/data-and-environment-curation-for-post-training-llms",
    "videoId": "ewtOo0scUh0",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1152000,
    "speakers": [
      {
        "slug": "mahesh-sathiamoorthy",
        "name": "Mahesh Sathiamoorthy",
        "organization": "Bespoke Labs"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Bespoke Labs CEO Mahesh Sathiamoorthy explains why high-quality synthetic data and reinforcement-learning environments are the main bottlenecks in post-training reliable, autonomous LLM agents. He traces Bespoke Curator and Bespoke Stratos into the collaborative OpenThoughts reasoning-data project, describes question selection, filtering, answer generation, and scaling, and discusses agent benchmarks, enterprise deployment, sandbox infrastructure, and prompt-and-harness optimization."
  },
  {
    "slug": "data-is-your-differentiator-building-secure-and-tailored-ai-systems",
    "title": "Data is Your Differentiator: Building Secure and Tailored AI Systems",
    "url": "https://ai.engineer/talks/data-is-your-differentiator-building-secure-and-tailored-ai-systems",
    "videoId": "ROfHHJmumcc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1210000,
    "speakers": [
      {
        "slug": "mani-khanuja",
        "name": "Mani Khanuja",
        "organization": "Amazon Web Services (AWS)"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "AWS speaker Mani Khanuja explains how proprietary enterprise data differentiates secure, tailored generative AI applications. She outlines contextual chatbot and RAG architectures using Amazon Bedrock Data Automation, Amazon Bedrock Knowledge Bases, the Retrieve and RetrieveAndGenerate APIs, reranking and query decomposition, and Amazon Bedrock Guardrails for responsible responses and PII protection. The talk closes by emphasizing testing, optimization, and production readiness."
  },
  {
    "slug": "data-quality-is-the-compute-multiplier",
    "title": "Data Quality is the Compute Multiplier",
    "url": "https://ai.engineer/talks/data-quality-is-the-compute-multiplier",
    "videoId": "_PdK6x7PQNM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1145000,
    "speakers": [
      {
        "slug": "ari-morcos",
        "name": "Ari Morcos",
        "organization": "DatologyAI"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "DatologyAI co-founder and CEO Ari Morcos argues that better training data multiplies the value of scarce compute. He outlines data mixing, quality-weighted sampling, task-distribution matching, and curriculum design; connects data pruning with improved neural scaling laws; and presents vision-language-model, multilingual, legal-domain, and open-model examples in which targeted curation improves model quality or efficiency."
  },
  {
    "slug": "data-readiness-is-a-myth-reliable-ai-with-an-agentic-semantic-layer-anushrut-gupta-promptql",
    "title": "\"Data readiness\" is a Myth: Reliable AI with an Agentic Semantic Layer — Anushrut Gupta, PromptQL",
    "url": "https://ai.engineer/talks/data-readiness-is-a-myth-reliable-ai-with-an-agentic-semantic-layer-anushrut-gupta-promptql",
    "videoId": "1nOTQsfe1RU",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1022000,
    "speakers": [
      {
        "slug": "anushrut-gupta",
        "name": "Anushrut Gupta",
        "organization": "PromptQL"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "PromptQL applied AI lead Anushrut Gupta argues that waiting for perfectly prepared enterprise data makes reliable AI deployments impractical because schemas, business definitions, and organizational knowledge continually change. He critiques manually maintained semantic layers and knowledge graphs, then presents an agentic semantic layer in which an LLM generates PromptQL plans that a deterministic runtime executes across distributed data sources. Examples show how the system can discover relevant tables and adapt to company-specific terminology and analytical requirements."
  },
  {
    "slug": "decoding-mistral-ai-s-large-language-models",
    "title": "Decoding Mistral AI's Large Language Models",
    "url": "https://ai.engineer/talks/decoding-mistral-ai-s-large-language-models",
    "videoId": "AUuktOQPWYg",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1096000,
    "speakers": [
      {
        "slug": "devendra-chaplot",
        "name": "Devendra Chaplot",
        "organization": "Mistral"
      },
      {
        "slug": "devendra-singh-chaplot",
        "name": "Devendra Singh Chaplot",
        "organization": "Mistral AI"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Mistral AI researcher Devendra Singh Chaplot surveys Mistral 7B, the sparse mixture-of-experts Mixtral 8x7B and 8x22B models, and Codestral 22B. He explains why open model distribution can complement commercial offerings, outlines cloud and private deployment options, and describes LLM pre-training, instruction tuning, reinforcement learning from human feedback, and direct preference optimization. The presentation concludes with model-efficiency benchmarking and Codestral's fill-in-the-middle coding capabilities, API access, and VS Code and JetBrains integrations."
  },
  {
    "slug": "deepswe-a-contamination-resistant-coding-benchmark-james-shi-datacurve",
    "title": "DeepSWE: A Contamination-Resistant Coding Benchmark — James Shi, Datacurve",
    "url": "https://ai.engineer/talks/deepswe-a-contamination-resistant-coding-benchmark-james-shi-datacurve",
    "videoId": "Yk87oUPVaxU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1054000,
    "speakers": [
      {
        "slug": "james-shi",
        "name": "James Shi",
        "organization": "Datacurve"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Datacurve founding engineer James Shi presents DeepSWE, a contamination-resistant benchmark of 113 original, long-horizon software-engineering tasks spanning five programming languages. He contrasts its authored tasks, broader repository coverage, Git-history safeguards, and behavior-focused verifiers with weaknesses in pull-request-derived benchmarks, then examines leaderboard separation, Claude and GPT failure patterns, task adherence, self-testing, and evaluation reliability."
  },
  {
    "slug": "define-ai-engineer",
    "title": "#define AI Engineer",
    "url": "https://ai.engineer/talks/define-ai-engineer",
    "videoId": "avWhreBUYF0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2465000,
    "speakers": [
      {
        "slug": "greg-brockman",
        "name": "Greg Brockman",
        "organization": "OpenAI"
      },
      {
        "slug": "swyx",
        "name": "swyx"
      },
      {
        "slug": "jensen-huang",
        "name": "Jensen Huang",
        "organization": "NVIDIA"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "In a conference fireside chat, OpenAI's Greg Brockman discusses self-directed mathematics and programming, learning PHP through W3Schools, leaving Harvard and MIT, and joining early-stage Stripe. The discussion moves to machine-learning history, OpenAI's research-engineering relationship, Codex and developer productivity, long-horizon agent execution and VM checkpointing, and the data-center infrastructure and co-design required for AGI; the recording also features a remote infrastructure question associated with named guest Jensen Huang."
  },
  {
    "slug": "defying-gravity",
    "title": "Defying Gravity",
    "url": "https://ai.engineer/talks/defying-gravity",
    "videoId": "HN-F-OQe6j0",
    "event": "AI Engineer Code 2025",
    "durationMs": 1510000,
    "speakers": [
      {
        "slug": "kevin-hou",
        "name": "Kevin Hou",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Kevin Hou introduces Google Antigravity, Google DeepMind’s agent-first development platform built around an AI editor, an agent-controlled browser, and Agent Manager. He explains how Gemini 3-era model improvements enable autonomous browser interaction, application verification, asynchronous task orchestration, multimodal image generation, and visual feedback, while arguing that new model capabilities should drive new software-development interfaces."
  },
  {
    "slug": "design-like-karpathy-is-watching",
    "title": "Design like Karpathy is watching 😎",
    "url": "https://ai.engineer/talks/design-like-karpathy-is-watching",
    "videoId": "huQPkrwVWwc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1166000,
    "speakers": [
      {
        "slug": "zeke-sikelianos",
        "name": "Zeke Sikelianos",
        "organization": "Replicate"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Replicate's Zeke Sikelianos examines Andrej Karpathy's experience building MenuGen to show how outdated documentation, API changes, rate limiting, deployment friction, and account onboarding undermine AI-assisted development. He advocates LLM-readable documentation through llms.txt and Markdown, machine-readable API schemas and MCP integrations, accessible structured product information, and treating documentation as an essential part of shipping software. The recording concludes with audience questions."
  },
  {
    "slug": "designing-agents-the-floor-is-the-frontier",
    "title": "Designing Agents (The Floor Is the Frontier)",
    "url": "https://ai.engineer/talks/designing-agents-the-floor-is-the-frontier",
    "videoId": "jHMiYtjoJfA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1186000,
    "speakers": [
      {
        "slug": "ben-hylak",
        "name": "Ben Hylak",
        "organization": "Raindrop"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Raindrop CTO and co-founder Ben Hylak argues that production AI agents need a higher reliability floor more than impressive peak capability. He contrasts brittle tool-specific evaluations with production issue detection, fix verification, simulation, local code-aware testing, and Raindrop Workshop's open-source tracing and self-healing workflows. The talk also examines user oversight, harmful autonomous actions, temporal clustering, and MCP-related code mode."
  },
  {
    "slug": "designing-ai-intensive-applications",
    "title": "Designing AI-Intensive Applications",
    "url": "https://ai.engineer/talks/designing-ai-intensive-applications",
    "videoId": "IHkyFhU6JEY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 782000,
    "speakers": [
      {
        "slug": "swyx",
        "name": "Shawn Wang (swyx)",
        "organization": "Latent.Space"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "AI Engineer organizer Shawn Wang (swyx) opens the 2025 World's Fair by discussing conference growth, audience-selected tracks, MCP-enabled conference tools, and the evolution of AI engineering. He surveys simple agent scaffolds, SWE-bench, RAG, DSPy, and software-development workflows, then argues that increasingly AI-intensive applications should be assessed by valuable AI output relative to human input rather than by disputes over whether systems are workflows or agents."
  },
  {
    "slug": "designing-ai-to-scale-human-thought",
    "title": "Designing AI to Scale Human Thought",
    "url": "https://ai.engineer/talks/designing-ai-to-scale-human-thought",
    "videoId": "0tVu-V3_fFw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 744000,
    "speakers": [
      {
        "slug": "jun-yu-tan",
        "name": "Jun Yu Tan",
        "organization": "Tusk"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Tusk founding engineer Jun Yu Tan argues that AI interfaces should augment human judgment rather than automate people out of complex work. Using email composition and code review as examples, he describes human-in-the-loop interaction patterns, verified AI-generated tests, systematic pessimism for discovering edge cases, and a novelty-criticality framework for prioritizing useful feedback. He concludes that effective augmentation should build user capabilities and measure growth alongside conventional product engagement."
  },
  {
    "slug": "designing-voice-agents-for-real-conversations",
    "title": "Designing Voice Agents for Real Conversations",
    "url": "https://ai.engineer/talks/designing-voice-agents-for-real-conversations",
    "videoId": "hMlLw1LeIK8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1977000,
    "speakers": [
      {
        "slug": "chintan-agrawal",
        "name": "Chintan Agrawal",
        "organization": "Amazon Web Services (AWS)"
      },
      {
        "slug": "daniel-wirjo",
        "name": "Daniel Wirjo",
        "organization": "Amazon Web Services (AWS)"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "AWS solutions architects Chintan Agrawal and Daniel Wirjo explain why natural voice agents depend on low-latency audio engineering and accurate conversational turn-taking, not just the underlying language model. They compare three approaches: Silero voice activity detection, speech-provider endpointing from Cartesia and Deepgram, and locally controlled Smart Turn classification layered over VAD. The presentation also covers streaming STT–LLM–TTS pipelines, interruption cancellation, latency tradeoffs, and a live Pipecat-based travel-assistant demonstration using Claude Haiku."
  },
  {
    "slug": "develop-at-idea-velocity",
    "title": "Develop at Idea Velocity",
    "url": "https://ai.engineer/talks/develop-at-idea-velocity",
    "videoId": "9arM9b7JgOo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 928000,
    "speakers": [
      {
        "slug": "jeffrey-lee-chan",
        "name": "Jeffrey Lee-Chan",
        "organization": "Snapchat"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Jeffrey Lee-Chan presents an interactive OpenClaw workshop on developing with parallel coding agents. He describes persistent Slack-derived memory, isolated worktrees, tmux-managed terminals, orchestrator-and-worker architectures, and Claude Code subagents, emphasizing separation between task-level specifications and implementation context. Audience questions also touch on multi-model workflows, conversational applications, and Docker-based sandboxing."
  },
  {
    "slug": "developer-experience-in-the-age-of-ai-coding-agents",
    "title": "Developer Experience in the Age of AI Coding Agents",
    "url": "https://ai.engineer/talks/developer-experience-in-the-age-of-ai-coding-agents",
    "videoId": "rT2Del5pwg4",
    "event": "AI Engineer Code 2025",
    "durationMs": 1100000,
    "speakers": [
      {
        "slug": "max-kanat-alexander",
        "name": "Max Kanat-Alexander",
        "organization": "Capital One"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Max Kanat-Alexander presents durable developer-experience investments that improve productivity for both software engineers and AI coding agents: standardized development environments, agent-friendly CLIs and APIs, explicit system context, and scalable code-review practices. He explains how growing volumes of agent-generated pull requests require clearer ownership, distributed reviewer assignments, enforceable SLOs, and better review quality, concluding that improvements benefiting developers also strengthen agent effectiveness."
  },
  {
    "slug": "developing-taste-in-coding-agents-applied-meta-neuro-symbolic-rl-ahmad-awais-command-code",
    "title": "Developing Taste in Coding Agents: Applied Meta Neuro-Symbolic RL — Ahmad Awais, Command Code",
    "url": "https://ai.engineer/talks/developing-taste-in-coding-agents-applied-meta-neuro-symbolic-rl-ahmad-awais-command-code",
    "videoId": "kWOQS3XPZ10",
    "event": "AI Engineer Code 2025",
    "durationMs": 1252000,
    "speakers": [
      {
        "slug": "ahmad-awais",
        "name": "Ahmad Awais",
        "organization": "Langbase; Command Code"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Ahmad Awais, founder and CEO of Langbase, introduces Command Code, a coding agent designed to learn an individual developer’s evolving preferences. Comparing it with Claude Code, he demonstrates a personalized TypeScript CLI that automatically incorporates tsup, Commander, pnpm, Vitest, and preferred version-flag conventions. He describes persistent taste files, explicit and implicit feedback, and a meta neuro-symbolic reinforcement-learning approach intended to capture architectural choices beyond generic code generation."
  },
  {
    "slug": "devin-2-0-and-the-future-of-swe",
    "title": "Devin 2.0 and the Future of SWE",
    "url": "https://ai.engineer/talks/devin-2-0-and-the-future-of-swe",
    "videoId": "MI83buT_23o",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 973000,
    "speakers": [
      {
        "slug": "scott-wu",
        "name": "Scott Wu",
        "organization": "Cognition"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Scott Wu outlines Cognition’s view of rapidly increasing AI software-agent task horizons and Devin’s progression from reliable, repetitive code migrations to broader autonomous engineering. He explains Playbooks for step-by-step instruction following, remote environments for repository setup, linting and CI, and the role of codebase intelligence, DeepWiki, and Devin 2.0’s collaborative IDE experience."
  },
  {
    "slug": "dispatch-from-the-future-building-an-ai-native-company-dan-shipper-every-ai-i",
    "title": "Dispatch from the Future: building an AI-native Company – Dan Shipper, Every, AI & I",
    "url": "https://ai.engineer/talks/dispatch-from-the-future-building-an-ai-native-company-dan-shipper-every-ai-i",
    "videoId": "MGzymaYBiss",
    "event": "AI Engineer Code 2025",
    "durationMs": 1078000,
    "speakers": [
      {
        "slug": "dan-shipper",
        "name": "Dan Shipper",
        "organization": "Every"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Every CEO and cofounder Dan Shipper argues that universal engineering adoption of AI changes how small companies build software. Using Every’s Cora email assistant and Claude Code as examples, he describes agents producing most of the company’s code, enabling individual engineers to build sophisticated products, and reducing collaboration overhead. He outlines compounding engineering: feeding planning, delegation, assessment, and implementation knowledge into reusable project instructions, sub-agents, and commands so subsequent work becomes easier."
  },
  {
    "slug": "does-genai-belong-to-data-scientists",
    "title": "Does GenAI \"belong\" to data scientists?",
    "url": "https://ai.engineer/talks/does-genai-belong-to-data-scientists",
    "videoId": "NKwIX3CiRgU",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1134000,
    "speakers": [
      {
        "slug": "phil-hetzel",
        "name": "Phil Hetzel",
        "organization": "Braintrust"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Braintrust solutions engineering leader Phil Hetzel argues that generative AI and agent development should not belong exclusively to data scientists or machine-learning platform teams. Contrasting traditional enterprises with cross-functional AI-native teams, he explains that foundation-model providers already handle model training, shifting application development toward prompts, context, evaluation, observability, and production feedback informed by human judgment. Audience questions explore organizational boundaries and closing the evaluation loop."
  },
  {
    "slug": "domain-adaptation-and-fine-tuning-for-domain-specific-llms",
    "title": "Domain adaptation and fine-tuning for domain-specific LLMs",
    "url": "https://ai.engineer/talks/domain-adaptation-and-fine-tuning-for-domain-specific-llms",
    "videoId": "cXPYtkosXG4",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1509000,
    "speakers": [
      {
        "slug": "abi-aryan",
        "name": "Abi Aryan"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Abi Aryan surveys practical strategies for adapting large language models to specialized domains, comparing prompting, retrieval-augmented generation, instruction tuning, full-model adaptation, and parameter-efficient fine-tuning. The talk explains LoRA and four-bit QLoRA, discusses bitsandbytes compatibility and deployment tradeoffs, distinguishes adaptive and behavioral fine-tuning, and highlights gradient checkpointing as a memory-saving technique."
  },
  {
    "slug": "don-t-be-data-poor",
    "title": "Don't be data poor",
    "url": "https://ai.engineer/talks/don-t-be-data-poor",
    "videoId": "XAsb7MIAzm8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1006000,
    "speakers": [
      {
        "slug": "anuj-iravane",
        "name": "Anuj Iravane",
        "organization": "Anterior"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      }
    ],
    "summary": "Anterior Head of AI Anuj Iravane explains how contractual restrictions on retaining protected medical records complicate evaluation of high-stakes healthcare AI. He describes why one-shot LLM generation produces insufficiently diverse synthetic records, then outlines a policy-guided approach using symbolic decision trees, diverse scenario sampling, and coarse-to-fine record construction. The resulting workflow incorporates clinical expertise and can reduce expensive ground-truthing."
  },
  {
    "slug": "don-t-build-slop-4-levels-of-ai-agent-maturity",
    "title": "Don't Build Slop (4 Levels of AI Agent Maturity)",
    "url": "https://ai.engineer/talks/don-t-build-slop-4-levels-of-ai-agent-maturity",
    "videoId": "yUmS-F9IX90",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1132000,
    "speakers": [
      {
        "slug": "ara-khan",
        "name": "Ara Khan",
        "organization": "Cline"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Ara Khan of Cline outlines four levels of AI-agent maturity, progressing from framework-based prototypes to custom agents, Kanban-oriented workflows, and cloud deployment. He argues for simpler prompts and architectures, CLI-first evaluation pipelines, agent-readable project instructions and CI/CD, careful software design, and awareness of frontier-model API lock-in. An audience question explores how Kanban workflows handle planning, human input, and review-state transitions."
  },
  {
    "slug": "don-t-get-one-shotted-use-ai-to-test-review-merge-and-deploy-code-tomas-reimers-graphite",
    "title": "Don’t get one-shotted: Use AI to test, review, merge, and deploy code — Tomas Reimers, Graphite",
    "url": "https://ai.engineer/talks/don-t-get-one-shotted-use-ai-to-test-review-merge-and-deploy-code-tomas-reimers-graphite",
    "videoId": "H6MrR5NbTZA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 345000,
    "speakers": [
      {
        "slug": "tomas-reimers",
        "name": "Tomas Reimers",
        "organization": "Graphite"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Graphite co-founder Tomas Reimers argues that AI-assisted coding accelerates the development inner loop while turning code review, testing, merging, and deployment into an outer-loop bottleneck. He presents Diamond, Graphite’s then-current AI code reviewer, as part of an AI-native workflow combining contextual pull-request feedback, CI integration, merge queues, security safeguards, and suggested fixes. He reports approximately 52% acceptance for Diamond comments versus roughly 45–50% for human comments, with AI-comment downvotes below 4%; these figures are speaker-reported and not independently verified."
  },
  {
    "slug": "don-t-just-slap-on-a-chatbot-building-ai-that-works-before-you-ask",
    "title": "Don't just slap on a chatbot: building AI that works before you ask",
    "url": "https://ai.engineer/talks/don-t-just-slap-on-a-chatbot-building-ai-that-works-before-you-ask",
    "videoId": "2cEGQEllBGc",
    "event": "AI Engineer Summit 2025",
    "durationMs": 346000,
    "speakers": [
      {
        "slug": "arthur-objartel",
        "name": "Arthur Objartel",
        "organization": "Evil Martians"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Evil Martians product designer Arthur Objartel argues that AI should anticipate user needs within existing workflows instead of relying on reactive chatbot interfaces. Using the Tegon issue tracker, he demonstrates contextual suggestions, automatic issue decomposition, and combined question-and-action assistance. He emphasizes reversible interactions, user agency, noncoercive recommendations, and identifying workflow friction and behavioral patterns to design proactive assistance across professional tools."
  },
  {
    "slug": "don-t-let-the-llm-drive-ornella-bahidika-joel-allou-microsoft",
    "title": "Don't Let the LLM Drive - Ornella Bahidika & Joel Allou, Microsoft",
    "url": "https://ai.engineer/talks/don-t-let-the-llm-drive-ornella-bahidika-joel-allou-microsoft",
    "videoId": "m24UKZomm7k",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 368000,
    "speakers": [
      {
        "slug": "ornella-bahidika",
        "name": "Ornella Bahidika",
        "organization": "Microsoft"
      },
      {
        "slug": "joel-allou",
        "name": "Joel Allou"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Ornella Bahidika and Joel Allou present Ace, a live AI voice tutor whose lessons are controlled by an external harness rather than an LLM. A state machine governs introduction, teaching, checking, grading, advancement, and wrap-up; each step gives the model a narrow task while application code validates outputs and controls transitions. They explain how this prevents skipped steps, loops, and premature completion, supports harness-controlled whiteboard interactions, and enables the smaller Claude Haiku 4.5 model to perform effectively within a constrained workflow."
  },
  {
    "slug": "don-t-ship-skills-without-evals",
    "title": "Don't Ship Skills Without Evals",
    "url": "https://ai.engineer/talks/don-t-ship-skills-without-evals",
    "videoId": "0vphxNt4wyk",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1306000,
    "speakers": [
      {
        "slug": "philipp-schmid",
        "name": "Philipp Schmid",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Google DeepMind engineer Philipp Schmid argues that nondeterministic coding-agent skills need evaluation before release. He explains concise triggering directives, context-specific references, positive and negative test cases, and removal of ineffective instructions; demonstrates a lightweight evaluation harness for a Gemini Interactions API skill using trace and regex validators; and recommends with-versus-without-skill ablations to measure usefulness and determine retirement."
  },
  {
    "slug": "dream-machine-scaling-to-1m-users-in-4-days-keegan-mccallum-luma-ai",
    "title": "Dream Machine: Scaling to 1m users in 4 days — Keegan McCallum, Luma AI",
    "url": "https://ai.engineer/talks/dream-machine-scaling-to-1m-users-in-4-days-keegan-mccallum-luma-ai",
    "videoId": "EY4O9M6AsWI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1143000,
    "speakers": [
      {
        "slug": "keegan-mccallum",
        "name": "Keegan McCallum",
        "organization": "Luma AI"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Keegan McCallum recounts how Luma AI scaled Dream Machine to one million users in four days, expanding from roughly 500 to 5,000 H100 GPUs within six hours and drawing on approximately 4,000 additional training-cluster GPUs. He describes Luma's multimodal ambitions and Modify Video API, operational problems with Triton Inference Server across multiple GPUs, nodes, and chipsets, pull-based queue scheduling, service-tier waiting-time tradeoffs, and PyTorch-level optimization. Closing questions address Kubernetes, accelerator portability, and vision-language-model-assisted prompting."
  },
  {
    "slug": "dspy-the-end-of-prompt-engineering",
    "title": "DSPy: The End of Prompt Engineering",
    "url": "https://ai.engineer/talks/dspy-the-end-of-prompt-engineering",
    "videoId": "-cKUW6n8hBU",
    "event": "AI Engineer Code 2025",
    "durationMs": 4393000,
    "speakers": [
      {
        "slug": "kevin-madura",
        "name": "Kevin Madura",
        "organization": "AlixPartners"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "AlixPartners consultant Kevin Madura presents a hands-on DSPy workshop arguing that enterprise AI systems should be built as testable Python programs rather than collections of manually tuned prompts. He demonstrates typed outputs and adapters, multimodal document ingestion and SEC filing classification, model routing and token observability, visual document-boundary detection, and metric-driven prompt optimization with GEPA, before discussing a repository for optimized DSPy programs and answering audience questions."
  },
  {
    "slug": "e-values-evaluating-the-values-of-ai",
    "title": "E-Values: Evaluating the Values of AI",
    "url": "https://ai.engineer/talks/e-values-evaluating-the-values-of-ai",
    "videoId": "mpzktdYTcdE",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1693000,
    "speakers": [
      {
        "slug": "sheila-gulati",
        "name": "Sheila Gulati",
        "organization": "Tola Capital"
      },
      {
        "slug": "nischal-nadhamuni",
        "name": "Nischal Nadhamuni",
        "organization": "Klarity"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Tola Capital founder Sheila Gulati and Klarity co-founder and CTO Nischal Nadhamuni examine why increasingly agentic AI systems require evaluations grounded in operational objectives, human values, and real-world performance. Gulati contrasts narrow task testing with more adaptive benchmarks, discusses black-box models and Transformer-based LLMs, and situates evaluation within the changing AI investment landscape. Nadhamuni explains Klarity's automation of document-heavy finance and accounting workflows and the need for customer-specific labels in bespoke enterprise AI systems."
  },
  {
    "slug": "effective-agent-design-patterns-in-production",
    "title": "Effective agent design patterns in production",
    "url": "https://ai.engineer/talks/effective-agent-design-patterns-in-production",
    "videoId": "72XxWkd8Jrk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 938000,
    "speakers": [
      {
        "slug": "laurie-voss",
        "name": "Laurie Voss",
        "organization": "LlamaIndex"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "LlamaIndex VP of Developer Relations Laurie Voss explains when production AI agents outperform simpler RAG or chatbot approaches, particularly for unstructured data and software-integrated workflows. He introduces LlamaIndex, LlamaParse, LlamaCloud, and LlamaHub; reviews embeddings, vector retrieval, and agent introspection; and applies Anthropic-inspired patterns including routing, parallelization, and aggregating outputs from multiple model executions. He closes by pointing to a workflow tutorial for building a deep-research agent."
  },
  {
    "slug": "effective-ai-agents-need-data-flywheels-not-the-next-biggest-llm-sylendran-arunagiri-nvidia",
    "title": "Effective AI Agents Need Data Flywheels, Not The Next Biggest LLM –  Sylendran Arunagiri, NVIDIA",
    "url": "https://ai.engineer/talks/effective-ai-agents-need-data-flywheels-not-the-next-biggest-llm-sylendran-arunagiri-nvidia",
    "videoId": "6lTxD_oUjXQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1001000,
    "speakers": [
      {
        "slug": "sylendran-arunagiri",
        "name": "Sylendran Arunagiri",
        "organization": "NVIDIA"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "NVIDIA’s Sylendran Arunagiri explains how agent data flywheels continuously capture production interactions and user feedback, curate ground truth, fine-tune models, evaluate quality, and redeploy improvements. He describes NeMo Curator, Customizer, Evaluator, Guardrails, Retriever, Data Store, and NVIDIA NIM, then examines NVIDIA’s internal NVinfo employee-support agent. In its routing case study, a fine-tuned 8B model reportedly matched a 70B model’s accuracy, illustrating how smaller specialized models can improve operational efficiency without relying on ever-larger LLMs."
  },
  {
    "slug": "efficient-reinforcement-learning",
    "title": "Efficient Reinforcement Learning",
    "url": "https://ai.engineer/talks/efficient-reinforcement-learning",
    "videoId": "o15AaYl7Wu0",
    "event": "AI Engineer Code 2025",
    "durationMs": 1219000,
    "speakers": [
      {
        "slug": "rhythm-garg",
        "name": "Rhythm Garg",
        "organization": "Applied Compute"
      },
      {
        "slug": "linden-li",
        "name": "Linden Li",
        "organization": "Applied Compute"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Applied Compute co-founders Rhythm Garg and Linden Li explain how reinforcement learning specializes language models for enterprise automation and continuous improvement. They show how synchronous training creates straggler-driven GPU underutilization, illustrate sampling variability with a Qwen-30B experiment, and describe first-principles systems modeling to allocate inference and training resources, improve throughput, reduce latency, and deliver predictably fast customer training runs."
  },
  {
    "slug": "embeddings-are-stunting-agents-how-codeium-breaks-through-the-ceiling-for-retrieval",
    "title": "Embeddings are Stunting Agents: How Codeium Breaks Through the Ceiling for Retrieval",
    "url": "https://ai.engineer/talks/embeddings-are-stunting-agents-how-codeium-breaks-through-the-ceiling-for-retrieval",
    "videoId": "DuZXbinJ4Uc",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1122000,
    "speakers": [
      {
        "slug": "kevin-hou",
        "name": "Kevin Hou",
        "organization": "Windsurf"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Kevin Hou explains why AI coding agents need repository-specific context beyond conventional embeddings and ever-larger context windows. Using Codeium's IDE assistant, he examines retrieval-augmented code generation, long-context latency, embedding dimensionality, and misleading needle-in-a-haystack benchmarks, then describes pull-request-derived evaluations, production cost constraints, and fast relevance ranking across codebases."
  },
  {
    "slug": "emergence-launch-ai-agents-and-the-future-enterprise",
    "title": "Emergence Launch: AI Agents and the future enterprise",
    "url": "https://ai.engineer/talks/emergence-launch-ai-agents-and-the-future-enterprise",
    "videoId": "xhQfRSueFZI",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 341000,
    "speakers": [
      {
        "slug": "satya-nitta",
        "name": "Satya Nitta",
        "organization": "Emergence"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Satya Nitta introduces Emergence’s enterprise-agent strategy and its Orchestrator, which coordinates LLMs and specialized agents to plan, verify, and automate complex business workflows. He presents Agent-E, an open-source browser automation agent evaluated on WebVoyager, and outlines research into agent self-improvement, planning, reasoning, and agent-oriented programming."
  },
  {
    "slug": "emulated-the-data-for-fully-autonomous-software-engineers-and-companies",
    "title": "Emulated: The data for fully autonomous software engineers and companies",
    "url": "https://ai.engineer/talks/emulated-the-data-for-fully-autonomous-software-engineers-and-companies",
    "videoId": "zkX03APVj0M",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 993000,
    "speakers": [
      {
        "slug": "joseph-wang",
        "name": "Joseph Wang",
        "organization": "Emulated"
      },
      {
        "slug": "sid",
        "name": "Sid",
        "organization": "Emulated"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Joseph Wang and Emulated co-founder Sid argue that autonomous software-engineering agents need training environments extending beyond repository-only coding benchmarks. They describe simulating company workflows, production incidents, distributed clusters, live traffic, and repeated deployments, while explaining why single-container sandboxes cannot faithfully reproduce cloud resource provisioning, network partitions, operational telemetry, and other long-horizon infrastructure responsibilities."
  },
  {
    "slug": "ending-ai-slop",
    "title": "Ending AI Slop",
    "url": "https://ai.engineer/talks/ending-ai-slop",
    "videoId": "lCBf9slCanI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 990000,
    "speakers": [
      {
        "slug": "thais-castello-branco",
        "name": "Thais Castello Branco",
        "organization": "Taste Labs"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Taste Labs founder Thais Castello Branco explains how to reduce generic AI-generated creative work by decomposing subjective qualities into measurable, context-specific signals. Using Reducto's brand as an example, she describes converting design attributes into verifiable reinforcement-learning tasks, outlines limitations of LLM-as-a-judge and reward hacking, and argues for expert-curated preference data, careful human quality assurance, and preserving legitimate differences in taste."
  },
  {
    "slug": "engineering-better-evals-scalable-llm-evaluation-pipelines-that-work",
    "title": "Engineering Better Evals: Scalable LLM Evaluation Pipelines That Work",
    "url": "https://ai.engineer/talks/engineering-better-evals-scalable-llm-evaluation-pipelines-that-work",
    "videoId": "spvXj9tnWAQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1486000,
    "speakers": [
      {
        "slug": "dat-ngo",
        "name": "Dat Ngo",
        "organization": "Arize"
      },
      {
        "slug": "aman-khan",
        "name": "Aman Khan",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Arize AI presenter Dat Ngo explains how production teams combine observability, traces, LLM-as-a-judge, human feedback, and lower-cost evaluation signals to assess LLM applications and agents at scale. Examples involving Duolingo and Booking.com's Trip Planner illustrate evaluation volume, conversational failure analysis, and practical trade-offs among coverage, cost, latency, and embedding-based guardrails. Audience questions address retrospective versus real-time evaluation and confidence scoring. Aman Khan appears in the recording metadata, but his participation is not confirmed by the supplied transcript excerpts."
  },
  {
    "slug": "engineering-voice-agents-latency-quality-and-scale",
    "title": "Engineering voice agents: Latency, quality, and scale",
    "url": "https://ai.engineer/talks/engineering-voice-agents-latency-quality-and-scale",
    "videoId": "N7b1PJc7SFc",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1475000,
    "speakers": [
      {
        "slug": "rishabh-bhargava",
        "name": "Rishabh Bhargava",
        "organization": "Together AI"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Together AI voice-team leader Rishabh Bhargava explains how to engineer production voice agents around tight conversational latency budgets. He describes orchestrating streaming speech-to-text, LLM reasoning and tool calls, and text-to-speech; targeting approximately 100 ms P90 transcription completion; improving turn detection and streaming beyond Whisper-style chunking; and colocating models to reduce network overhead. Audience questions address function-calling evaluations, deployment topology, classifier-based guardrails, and routing."
  },
  {
    "slug": "ensure-ai-agents-work-evaluation-frameworks-for-scaling-success",
    "title": "Ensure AI Agents Work: Evaluation Frameworks for Scaling Success",
    "url": "https://ai.engineer/talks/ensure-ai-agents-work-evaluation-frameworks-for-scaling-success",
    "videoId": "OC04sP_QgTI",
    "event": "AI Engineer Summit 2025",
    "durationMs": 928000,
    "speakers": [
      {
        "slug": "aparna-dhinakaran",
        "name": "Aparna Dhinakaran",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Arize AI co-founder and Chief Product Officer Aparna Dhinakaran explains how to make production AI agents reliable through evaluation and observability across routers, skills, memory, tool calls, and execution traces. Examples include Priceline's Penny travel assistant, SQL-based trace investigation, RAG relevance and answer-correctness checks, LLM-as-a-judge evaluations, execution-path variability across model providers, and voice-agent metrics including sentiment, speech quality, and speech-to-text accuracy."
  },
  {
    "slug": "enterprise-agents-have-a-structure-problem-ishita-daga-tesla",
    "title": "Enterprise Agents Have a Structure Problem - Ishita Daga, Tesla",
    "url": "https://ai.engineer/talks/enterprise-agents-have-a-structure-problem-ishita-daga-tesla",
    "videoId": "B8l81jhvHbI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 728000,
    "speakers": [
      {
        "slug": "ishita-daga",
        "name": "Ishita Daga",
        "organization": "Tesla"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Tesla machine learning engineer Ishita Daga argues that enterprise data agents fail primarily because business context is ambiguous, stale, or dependent on individual and team preferences—not because models are too small. She recommends curated semantic layers containing authoritative KPI and business definitions, continuously updated enterprise data sources, and feedback loops, while identifying personalized metric preferences as an unresolved challenge beyond basic agent memory."
  },
  {
    "slug": "enterprise-deep-research-the-next-killer-app-for-enterprise-ai-ofer-mendelevitch-vectara",
    "title": "Enterprise Deep Research: The Next Killer App for Enterprise AI — Ofer Mendelevitch, Vectara",
    "url": "https://ai.engineer/talks/enterprise-deep-research-the-next-killer-app-for-enterprise-ai-ofer-mendelevitch-vectara",
    "videoId": "fh9LgKXBGnQ",
    "event": "AI Engineer Code 2025",
    "durationMs": 319000,
    "speakers": [
      {
        "slug": "ofer-mendelevitch",
        "name": "Ofer Mendelevitch",
        "organization": "Vectara"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Ofer Mendelevitch explains how Vectara adapts autonomous, multi-step deep research to private enterprise data using agentic RAG, multimodal ingestion, hybrid retrieval, parallel agents, and hallucination detection with HHEM. He outlines deployment and security capabilities and illustrates applications including automated RFP responses and onboarding guides assembled from enterprise documentation systems."
  },
  {
    "slug": "evaling-video-slop",
    "title": "Evaling Video Slop",
    "url": "https://ai.engineer/talks/evaling-video-slop",
    "videoId": "b_PmGocP4rc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1393000,
    "speakers": [
      {
        "slug": "maor-bril",
        "name": "Maor Bril",
        "organization": "Character.ai"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Character.ai's Maor Bril explains why generated video cannot be evaluated reliably using single-frame metrics, absolute quality scores, or uncalibrated AI judges. He describes JudgeJudy's combination of automated metrics, human annotation, pairwise preference comparisons, and a compact Qwen-based vision-language judge to detect temporal, visual, and audio failures earlier. Audience questions address sound assessment, model selection, scale, and OpenTelemetry instrumentation."
  },
  {
    "slug": "evals-101-doug-guthrie-braintrust",
    "title": "Evals 101 — Doug Guthrie, Braintrust",
    "url": "https://ai.engineer/talks/evals-101-doug-guthrie-braintrust",
    "videoId": "bk0TmxoZlUY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2911000,
    "speakers": [
      {
        "slug": "doug-guthrie",
        "name": "Doug Guthrie",
        "organization": "Braintrust"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Braintrust solutions engineer Doug Guthrie introduces practical AI evaluations, covering evaluation components, baseline quality measurement, SDK and platform workflows, custom scoring, agentic applications, production logging, online monitoring, and human feedback. Audience questions address production model comparison, LLM-as-a-judge scoring, and CI/CD integration."
  },
  {
    "slug": "evals-are-broken-use-them-anyway",
    "title": "Evals Are Broken, Use Them Anyway",
    "url": "https://ai.engineer/talks/evals-are-broken-use-them-anyway",
    "videoId": "QuuIywMG4s8",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1144000,
    "speakers": [
      {
        "slug": "ara-khan",
        "name": "Ara Khan",
        "organization": "Cline"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Cline's Ara Khan argues that coding-agent evaluations are indispensable but misleading when benchmark scores are treated as objective truth or replaced entirely by subjective impressions. He examines SWE-bench Verified's limitations, describes constructing evaluations that better reflect real programming workflows, and uses Terminal-Bench to show how standardized Linux environments, CPU and memory configuration, and model-specific agent harnesses affect results."
  },
  {
    "slug": "evals-are-not-unit-tests",
    "title": "Evals Are Not Unit Tests",
    "url": "https://ai.engineer/talks/evals-are-not-unit-tests",
    "videoId": "L8OoYeDI_ls",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 922000,
    "speakers": [
      {
        "slug": "ido-pesok",
        "name": "Ido Pesok",
        "organization": "Vercel"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Vercel engineer Ido Pesok explains why nondeterministic LLM applications need systematic evaluations rather than a handful of successful manual tests. Using a fruit-letter-counting application built with v0, AI SDK, and GPT-4.1, he shows how unexpected production inputs expose failures, then recommends collecting representative user examples, maintaining reusable evaluation data, and running Braintrust evaluations in CI to detect improvements and regressions."
  },
  {
    "slug": "evals-driven-development-engineering-a-mental-health-ai-coach-ethically-safely",
    "title": "Evals Driven-Development: Engineering a Mental Health AI Coach Ethically & Safely",
    "url": "https://ai.engineer/talks/evals-driven-development-engineering-a-mental-health-ai-coach-ethically-safely",
    "videoId": "O72p-rBb2bA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1277004,
    "speakers": [
      {
        "slug": "akele-reed",
        "name": "Akele Reed",
        "organization": "SonderMind"
      },
      {
        "slug": "dave-revere",
        "name": "Dave Revere",
        "organization": "SonderMind"
      },
      {
        "slug": "doug-keller",
        "name": "Doug Keller",
        "organization": "SonderMind"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Akele Reed and Dave Revere explain how SonderMind builds Sonder, a clinically grounded AI mental-health coach designed to complement therapy and direct users to human care when appropriate. They describe modular input and output guardrails around Sonder Core, independent LLM-as-a-judge safety checks, context-sensitive crisis intervention, and clinician-informed evaluations that capture risks generic moderation systems miss. The presentation closes with audience questions about model choices and calibration."
  },
  {
    "slug": "evals-workshop-mastering-ai-evaluation-from-playground-to-production",
    "title": "[Evals Workshop] Mastering AI Evaluation: From Playground to Production",
    "url": "https://ai.engineer/talks/evals-workshop-mastering-ai-evaluation-from-playground-to-production",
    "videoId": "9iN-cPnp7xg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 5108000,
    "speakers": [
      {
        "slug": "carlos-esteban",
        "name": "Carlos Esteban",
        "organization": "Braintrust"
      },
      {
        "slug": "doug",
        "name": "Doug",
        "organization": "Braintrust"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Braintrust solutions engineers Carlos Esteban and Doug lead a hands-on AI evaluation workshop covering platform and SDK workflows, longitudinal experiments, code-based scoring, OpenAI configuration, TypeScript evaluation files, CI integration, prompt loading, and human review. Audience discussions examine confidence in LLM-as-a-judge results, traditional machine-learning alternatives, and subject-matter-expert review."
  },
  {
    "slug": "evaluating-ai-search-a-practical-framework-for-augmented-ai-systems",
    "title": "Evaluating AI Search: A Practical Framework for Augmented AI Systems",
    "url": "https://ai.engineer/talks/evaluating-ai-search-a-practical-framework-for-augmented-ai-systems",
    "videoId": "wRJD0inpmjU",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1233000,
    "speakers": [
      {
        "slug": "julia-neagu",
        "name": "Julia Neagu",
        "organization": "Quotient AI"
      },
      {
        "slug": "deanna-emery",
        "name": "Deanna Emery",
        "organization": "Quotient AI"
      },
      {
        "slug": "maitar-asher",
        "name": "Maitar Asher",
        "organization": "Tavily"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Julia Neagu and Deanna Emery of Quotient AI join Maitar Asher of Tavily to explain how production AI-search agents can be evaluated despite changing web content and unpredictable user queries. They contrast static SimpleQA and HotpotQA benchmarks with dynamically generated, multi-source evaluation datasets, describe a LangGraph-based generation workflow and LangSmith experiment tracking, and demonstrate reference-free metrics such as answer completeness alongside broader hallucination and observability concerns."
  },
  {
    "slug": "events-are-the-wrong-abstraction-for-your-ai-agents",
    "title": "Events are the Wrong Abstraction for Your AI Agents",
    "url": "https://ai.engineer/talks/events-are-the-wrong-abstraction-for-your-ai-agents",
    "videoId": "KJ9eZYTWS1Y",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 880000,
    "speakers": [
      {
        "slug": "mason-egger",
        "name": "Mason Egger",
        "organization": "Temporal Technologies"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Mason Egger of Temporal argues that AI agents inherit familiar distributed-systems problems and that event-driven architectures introduce fragmented APIs, queue-management overhead, ad hoc state machines, race conditions, and misleading assumptions about loose coupling. He presents durable execution and Temporal's MIT-licensed platform as a simpler, resilient abstraction that lets developers concentrate on application logic."
  },
  {
    "slug": "every-company-should-have-a-brain-garry-tan-y-combinator",
    "title": "Every company should have a Brain — Garry Tan, Y Combinator",
    "url": "https://ai.engineer/talks/every-company-should-have-a-brain-garry-tan-y-combinator",
    "videoId": "eBUyTS7SzV4",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1268000,
    "speakers": [
      {
        "slug": "garry-tan",
        "name": "Garry Tan",
        "organization": "Y Combinator"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Y Combinator president and CEO Garry Tan argues that AI-native companies gain leverage by encoding organizational knowledge, responsibilities, and reusable procedures as markdown-based agent skills rather than treating AI as autocomplete. Drawing on his reported productivity gains and examples from YC, he describes resolver tables, tests as agent performance reviews, lean startup operations, and the need to combine deterministic computation with model capabilities. He presents GBrain and the broader company-brain concept as systems for retrieving institutional knowledge and providing agents with the right context."
  },
  {
    "slug": "every-harness-will-become-a-claw",
    "title": "Every Harness Will Become A Claw",
    "url": "https://ai.engineer/talks/every-harness-will-become-a-claw",
    "videoId": "8qWIPUia2O8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 936000,
    "speakers": [
      {
        "slug": "sam-bhagwat",
        "name": "Sam Bhagwat",
        "organization": "Mastra"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Mastra founder and CEO Sam Bhagwat argues that AI systems are progressing from LLMs and agent loops to sophisticated coding-agent harnesses and ultimately always-on personal agents, or claws. He describes harness features including planning, parallel subagents, skills, background execution, persistent threads, and interactive steering, then explains how cloud sandboxes, Slack collaboration, messaging channels, and durable memory extend these systems into continuously available agents. He discusses OpenClaw and Hermes Agent as examples and predicts eventual consolidation across the expanding agent ecosystem."
  },
  {
    "slug": "every-solo-agent-builder-eventually-reinvents-a-worse-version-of-ci-cd",
    "title": "Every Solo Agent Builder Eventually Reinvents a Worse Version of CI/CD",
    "url": "https://ai.engineer/talks/every-solo-agent-builder-eventually-reinvents-a-worse-version-of-ci-cd",
    "videoId": "WLXxTaPagA8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 651000,
    "speakers": [
      {
        "slug": "sumaiya-shrabony",
        "name": "Sumaiya Shrabony"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Sumaiya Shrabony examines her open-source, 19-skill Claude Code content system to show how solo agent builders inadvertently recreate CI/CD capabilities, including regression testing, monitoring, contract testing, staging, and audit trails. Using seven pipeline handoffs, she demonstrates three deceptively polished failure modes—voice drift, unverified claims, and duplicate content—and argues that explicit, blocking validation gates with audit records are essential before agent-generated artifacts reach production."
  },
  {
    "slug": "everything-i-learned-training-frontier-small-models",
    "title": "Everything I Learned Training Frontier Small Models",
    "url": "https://ai.engineer/talks/everything-i-learned-training-frontier-small-models",
    "videoId": "fLUtUkqYHnQ",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1213000,
    "speakers": [
      {
        "slug": "maxime-labonne",
        "name": "Maxime Labonne",
        "organization": "Liquid AI"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Liquid AI’s Maxime Labonne explains how to train and deploy frontier small language models for constrained edge devices. He contrasts Gemma and Qwen embedding overhead with LFM2’s gated-short-convolution and grouped-query-attention architecture, discusses CPU, smartphone, and GPU inference, and outlines an LFM2.5 training pipeline spanning 28 trillion tokens, supervised fine-tuning, preference alignment, and reinforcement learning. He also examines scaling behavior, reasoning-model doom loops, long-context workarounds using Python, and practical choices between small and large models."
  },
  {
    "slug": "everything-is-a-rollout-alex-shaw-ryan-marten-terminal-bench-harbor-laude-institute",
    "title": "Everything Is a Rollout — Alex Shaw + Ryan Marten, Terminal-Bench, Harbor, Laude Institute",
    "url": "https://ai.engineer/talks/everything-is-a-rollout-alex-shaw-ryan-marten-terminal-bench-harbor-laude-institute",
    "videoId": "jRCpXUjz4CI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1271000,
    "speakers": [
      {
        "slug": "alex-shaw",
        "name": "Alex Shaw",
        "organization": "Laude Institute"
      },
      {
        "slug": "ryan-marten",
        "name": "Ryan Marten",
        "organization": "Laude Institute"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Alex Shaw of Laude Institute argues that AI agents should be treated as black-box machine-learning systems whose behavior is improved through empirical evaluation and rollout-driven optimization. He introduces Harbor as infrastructure for sandboxed agent evaluations and reinforcement-learning environments, demonstrates parallel Terminal-Bench execution with Codex and GPT-5.5, and describes agentic MapReduce workflows for distributed execution, trajectory analysis, reward-hacking detection, and iterative product improvement. Ryan Marten is named in the recording metadata and candidate schedule, but his speaking participation is not supported by the supplied transcript excerpts."
  },
  {
    "slug": "everything-is-ugly-so-go-build-something-that-isn-t",
    "title": "Everything is ugly, so go build something that isn't",
    "url": "https://ai.engineer/talks/everything-is-ugly-so-go-build-something-that-isn-t",
    "videoId": "yG5d5UaGz1M",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1515000,
    "speakers": [
      {
        "slug": "raiza-martin",
        "name": "Raiza Martin",
        "organization": "Huxe"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Huxe founder and former Google NotebookLM product leader Raiza Martin argues that AI is collapsing traditional distinctions between product, engineering, and design while creating opportunities to rebuild interfaces around genuine user needs. Drawing on NotebookLM's origins as Project Tailwind, she advocates product purpose, restraint, clear user outcomes, and delight over novelty-driven AI demos or merely exposing model capabilities."
  },
  {
    "slug": "everything-we-knew-about-software-has-changed-theo-browne",
    "title": "Everything we knew about software has changed — Theo Browne",
    "url": "https://ai.engineer/talks/everything-we-knew-about-software-has-changed-theo-browne",
    "videoId": "xUnRQ9vLXxo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 962000,
    "speakers": [
      {
        "slug": "theo-browne",
        "name": "Theo Browne",
        "organization": "Ping Labs"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Theo Browne argues that increasingly capable AI coding models, from Claude 3.5 Sonnet and Claude Opus 4.5 to Mythos, invalidate assumptions about how software should be built and how ambitious small teams can be. Drawing on legacy terminal workflows, Apple's iOS 7 redesign, pull-request sunk costs, his Y Combinator-backed Ping startup, and automated GitHub repository review, he urges engineers to abandon inherited constraints and pursue greater product breadth and ambition."
  },
  {
    "slug": "everything-you-need-to-know-about-agent-observability",
    "title": "Everything You Need To Know About Agent Observability",
    "url": "https://ai.engineer/talks/everything-you-need-to-know-about-agent-observability",
    "videoId": "-aM2EDTiaMs",
    "event": "AI Engineer Europe 2026",
    "durationMs": 3025000,
    "speakers": [
      {
        "slug": "danny-gollapalli",
        "name": "Danny Gollapalli",
        "organization": "Raindrop"
      },
      {
        "slug": "ben-hylak",
        "name": "Ben Hylak",
        "organization": "Raindrop"
      },
      {
        "slug": "zubin-koticha",
        "name": "Zubin Koticha",
        "organization": "Raindrop"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Raindrop presenters Zubin Koticha and Danny Gollapalli explain how to monitor production AI agents using explicit telemetry and implicit signals such as refusals, task failures, user frustration, and moderation concerns. They discuss regex and classifier-based detection, release experiments and statistical relevance, then move into a hands-on coding-agent workshop covering instrumentation, self-diagnostics, trace analysis, and tool failures. Audience questions address experiment design, autonomous workflows, observability integrations, and trace volume."
  },
  {
    "slug": "everything-you-need-to-know-about-finetuning-and-merging-llms",
    "title": "Everything you need to know about Finetuning and Merging LLMs",
    "url": "https://ai.engineer/talks/everything-you-need-to-know-about-finetuning-and-merging-llms",
    "videoId": "uLrOI65XbDw",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1072000,
    "speakers": [
      {
        "slug": "maxime-labonne",
        "name": "Maxime Labonne",
        "organization": "Liquid AI"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Liquid AI's Maxime Labonne surveys practical LLM post-training and model merging, explaining why organizations fine-tune open models and comparing Unsloth, Hugging Face TRL, Axolotl, and LlamaFactory. He covers supervised fine-tuning, synthetic dataset construction, preference optimization with PPO, DPO, KTO, and IPO, dataset diversity and quality filtering, adapter-based training, four-bit quantization, merged-model evaluation, and open-source learning notebooks."
  },
  {
    "slug": "evolving-claude-apis-for-agents",
    "title": "Evolving Claude APIs for Agents",
    "url": "https://ai.engineer/talks/evolving-claude-apis-for-agents",
    "videoId": "aqW68Is_Kj4",
    "event": "AI Engineer Code 2025",
    "durationMs": 805000,
    "speakers": [
      {
        "slug": "katelyn-lesse",
        "name": "Katelyn Lesse",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Anthropic platform leader Katelyn Lesse explains how the Claude Developer Platform supports high-performing, long-running agents, using Claude Code as an example. She covers configurable reasoning, built-in and custom tools, MCP integrations, external memory, context editing, secure sandboxed code execution, and combining skills with tools and retrieved context."
  },
  {
    "slug": "exposing-agents-as-mcp-servers-with-mcp-agent-sarmad-qadri",
    "title": "Exposing Agents as MCP Servers with mcp-agent: Sarmad Qadri",
    "url": "https://ai.engineer/talks/exposing-agents-as-mcp-servers-with-mcp-agent-sarmad-qadri",
    "videoId": "uFPAtKIN-FQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1085000,
    "speakers": [
      {
        "slug": "sarmad-qadri",
        "name": "Sarmad Qadri",
        "organization": "LastMile AI"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Sarmad Qadri, identified in the recording as CEO of LastMile AI, explains how mcp-agent exposes AI agents as Model Context Protocol servers that compatible clients can invoke and compose. Drawing parallels with Language Server Protocol, he describes standardized access to tools and data, simpler architectures based on stronger models and Anthropic's Building Effective Agents patterns, orchestrator-based multi-agent systems, and durable asynchronous workflows that can pause, resume, retry, and incorporate human input."
  },
  {
    "slug": "eyelevel-launch-your-rag-is-tripping-here-s-the-real-reason-why",
    "title": "EyeLevel Launch: Your RAG is Tripping, Here's the Real Reason Why",
    "url": "https://ai.engineer/talks/eyelevel-launch-your-rag-is-tripping-here-s-the-real-reason-why",
    "videoId": "8RbTZl7bs5U",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 373000,
    "speakers": [
      {
        "slug": "benjamin-fletcher",
        "name": "Benjamin Fletcher",
        "organization": "EyeLevel.ai"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "EyeLevel.ai co-founder Benjamin Fletcher argues that enterprise RAG errors often originate in poor document ingestion, lost contextual information, and irrelevant retrieval rather than the language model itself. He describes an approach using fine-tuned vision models, multimodal extraction, semantic objects, multifield search, and LLM reranking instead of conventional vector-database retrieval, illustrated by an Air France customer-support copilot. Reported accuracy figures are company and speaker claims, not independent benchmark verification."
  },
  {
    "slug": "fast-models-need-slow-developers",
    "title": "Fast Models Need Slow Developers",
    "url": "https://ai.engineer/talks/fast-models-need-slow-developers",
    "videoId": "TeGsFFNqRLA",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1082000,
    "speakers": [
      {
        "slug": "sarah-chieng",
        "name": "Sarah Chieng",
        "organization": "Cerebras"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Cerebras developer-experience lead Sarah Chieng argues that ultra-fast coding models amplify both developer productivity and technical debt, requiring more deliberate human oversight. She explains how memory bandwidth, on-chip SRAM, KV caches, separated prefill/decode workloads, expert pruning, and inference-serving optimizations improve speed, then recommends pairing larger planning models with fast execution models, actively steering agents, and preserving context in external progress files."
  },
  {
    "slug": "feedback-loops-are-all-you-need",
    "title": "Feedback Loops are All You Need",
    "url": "https://ai.engineer/talks/feedback-loops-are-all-you-need",
    "videoId": "ON5LIT0M4do",
    "event": "AI Engineer Europe 2026",
    "durationMs": 611000,
    "speakers": [
      {
        "slug": "mehedi-hassan",
        "name": "Mehedi Hassan",
        "organization": "Granola"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Granola product engineer Mehedi Hassan explains why shipping AI meeting-chat features requires continuous feedback loops rather than one-shot implementation. He describes challenges with web search and role-specific outputs, then shows how internal tracing exposes tool calls, reasoning, search trails, and costs across teams. He also discusses Electron application architecture, pull-request preview links, and Cursor-assisted testing to accelerate experimentation and verification."
  },
  {
    "slug": "fighting-slop-with-slop",
    "title": "fighting slop with slop",
    "url": "https://ai.engineer/talks/fighting-slop-with-slop",
    "videoId": "AMiyLItEtLA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1292000,
    "speakers": [
      {
        "slug": "vaibhav-gupta",
        "name": "Vaibhav Gupta",
        "organization": "Boundary"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Boundary's Vaibhav Gupta describes building BAML with AI-assisted engineering practices that replace conventional code reviews with shared architecture documentation, versioned design documents, Slack-integrated workflows, CLI-enforced architectural invariants, and automated agent evaluations. He explains how generated BAML programs, transcript inspection, tool-call metrics, A/B tests, and first-principles tracing can guide the design of a programming language optimized for both humans and agents."
  },
  {
    "slug": "first-steps-toward-automated-ai-research",
    "title": "First Steps Toward Automated AI Research",
    "url": "https://ai.engineer/talks/first-steps-toward-automated-ai-research",
    "videoId": "pWXUkLP9uWM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1224000,
    "speakers": [
      {
        "slug": "richard-socher",
        "name": "Richard Socher",
        "organization": "You.com / Recursive Superintelligence"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Richard Socher outlines an evolution-inspired Eureka machine for automating scientific discovery and ultimately enabling recursively self-improving AI. He connects open-ended innovation with competing hypotheses, empirical testing, scientific data, simulations, physical experiments, and AI-oriented search infrastructure. He distinguishes genuinely self-improving research systems from short nanochat/autoresearch optimization runs and describes the broader pursuit of superintelligence."
  },
  {
    "slug": "five-hard-earned-lessons-about-evals",
    "title": "Five Hard-Earned Lessons About Evals",
    "url": "https://ai.engineer/talks/five-hard-earned-lessons-about-evals",
    "videoId": "a4BV0gGmXgA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1186000,
    "speakers": [
      {
        "slug": "ankur-goyal",
        "name": "Ankur Goyal",
        "organization": "Braintrust"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Braintrust founder Ankur Goyal explains five practical lessons for building AI evaluation systems: measure whether evals enable rapid model adoption, engineer realistic datasets and application-specific scorers, optimize agents' complete context and tool outputs, remain ready for model improvements, and improve datasets, prompts, and scoring together. Examples include Notion's 24-hour model turnaround, Braintrust's open-source Autoevals library, JSON-versus-YAML tool outputs, and the Loop evaluation agent. Audience questions address converting user feedback into evals and interpreting negative feedback."
  },
  {
    "slug": "fixing-bugs-in-gemma-llama-phi-3",
    "title": "Fixing bugs in Gemma, Llama & Phi-3",
    "url": "https://ai.engineer/talks/fixing-bugs-in-gemma-llama-phi-3",
    "videoId": "TKmfBnW0mQA",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1062000,
    "speakers": [
      {
        "slug": "daniel-han",
        "name": "Daniel Han",
        "organization": "Unsloth"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Unsloth's Daniel Han examines practical open-model fine-tuning failures, focusing on Llama 3 after briefly reviewing Gemma activation-function bugs and Phi-3 sliding-window and QKV issues. He explains how duplicated BOS tokens, incompatible base-versus-instruct chat templates, untrained token embeddings, and identical padding/EOS tokens can reduce accuracy, produce NaN gradients, or cause endless generation, then highlights Unsloth fixes, Colab notebooks, Ollama support, and LoRA tuning."
  },
  {
    "slug": "forget-rag-pipelines-build-production-ready-ai-agents-in-15-minutes",
    "title": "Forget RAG Pipelines—Build Production-Ready AI Agents in 15 Minutes",
    "url": "https://ai.engineer/talks/forget-rag-pipelines-build-production-ready-ai-agents-in-15-minutes",
    "videoId": "lArgRvBV3tQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 4543000,
    "speakers": [
      {
        "slug": "nina-lopatina",
        "name": "Nina Lopatina",
        "organization": "Contextual AI"
      },
      {
        "slug": "rajiv-shah",
        "name": "Rajiv Shah",
        "organization": "Contextual AI"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Nina Lopatina and Rajiv Shah lead a hands-on Contextual AI workshop on building managed, production-oriented RAG agents using a notebook and a single API key. They demonstrate document ingestion, agent creation through an API and graphical interface, financial and spurious-correlation examples, and natural-language evaluation with LMUnit. Audience questions address TypeScript and JavaScript integration, data sovereignty and residency, recent RAG developments, and model fine-tuning."
  },
  {
    "slug": "form-factors-for-your-new-ai-coworkers",
    "title": "Form factors for your new AI coworkers",
    "url": "https://ai.engineer/talks/form-factors-for-your-new-ai-coworkers",
    "videoId": "CiMVKnX-CNI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 935000,
    "speakers": [
      {
        "slug": "craig-wattrus",
        "name": "Craig Wattrus",
        "organization": "Flatfile"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Flatfile's Craig Wattrus explores AI coworkers as a design challenge, organizing interaction patterns into invisible, ambient, inline, and conversational experiences. Using personalized demos, agent-assisted data transformation, a v0-built prompt-tuning interface, and cursor-like collaborative agents, he argues that designers should shape agent character and experiment directly with AI capabilities rather than relying on static prototypes."
  },
  {
    "slug": "forward-deployed-engineering-101",
    "title": "Forward Deployed Engineering 101",
    "url": "https://ai.engineer/talks/forward-deployed-engineering-101",
    "videoId": "KwhgfwOSToQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1068000,
    "speakers": [
      {
        "slug": "kevin-bai",
        "name": "Kevin Bai",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Anthropic applied AI engineer Kevin Bai draws on his experience at Rippling and Palantir to explain forward-deployed engineering as an enterprise go-to-market model for selling complex technical platforms to nontechnical buyers. Using Palantir Foundry and its Ontology as examples, he contrasts reusable platform-based customer solutions with bespoke services, discusses enterprise contract value and design partnerships, and answers audience questions about shared primitives, collaboration between FDE teams, platform boundaries, and hiring customer-facing engineers."
  },
  {
    "slug": "forward-deployed-engineering-at-cursor-pauline-brunet",
    "title": "Forward Deployed Engineering at Cursor — Pauline Brunet",
    "url": "https://ai.engineer/talks/forward-deployed-engineering-at-cursor-pauline-brunet",
    "videoId": "APqXGyCoGW4",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1247000,
    "speakers": [
      {
        "slug": "pauline-brunet",
        "name": "Pauline Brunet",
        "organization": "Cursor"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Cursor forward deployed engineering leader Pauline Brunet explains how enterprise customer maturity and product customization determine whether embedded engineers, advisory support, or self-service are appropriate. She distinguishes FDE work from conventional SaaS deployment and staff augmentation, emphasizes clear customer outcomes and tightly scoped engagements, and discusses attracting and retaining engineers suited to collaborative, approximately six-week implementation cycles."
  },
  {
    "slug": "foundry-local-cutting-edge-ai-experiences-on-device-with-onnx-runtime-and-olive-emma-ning-micros",
    "title": "Foundry Local: Cutting-Edge AI Experiences on Device with ONNX Runtime and Olive — Emma Ning, Microsoft",
    "url": "https://ai.engineer/talks/foundry-local-cutting-edge-ai-experiences-on-device-with-onnx-runtime-and-olive-emma-ning-micros",
    "videoId": "l614N5W60ls",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1372000,
    "speakers": [
      {
        "slug": "emma-ning",
        "name": "Emma Ning",
        "organization": "Microsoft"
      },
      {
        "slug": "tsavo-knott",
        "name": "Tsavo Knott",
        "organization": "Pieces"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Microsoft product manager Emma Ning presents Foundry Local, an on-device AI stack for building cross-platform applications that remain responsive and usable offline while improving privacy. She explains its ONNX Runtime foundation, local model-management service, Azure AI Foundry integration, CLI and SDK, support for Windows and macOS, and acceleration across NVIDIA, Intel, AMD, and Qualcomm hardware. An embedded Pieces customer testimonial describes offline-first developer tooling, followed by local inference demonstrations and discussion of an agent-feature private preview."
  },
  {
    "slug": "from-46-to-90-fine-tuning-tiny-llms-for-on-device-agents-cormac-brick-google",
    "title": "From 46% to 90%: Fine-Tuning Tiny LLMs for On-Device Agents — Cormac Brick, Google",
    "url": "https://ai.engineer/talks/from-46-to-90-fine-tuning-tiny-llms-for-on-device-agents-cormac-brick-google",
    "videoId": "-TiET_K-E_g",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1261000,
    "speakers": [
      {
        "slug": "cormac-brick",
        "name": "Cormac Brick",
        "organization": "Google"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Google engineer Cormac Brick explains how to build on-device agents using tiny language models and the Google AI Edge stack. He contrasts system-provided Gemini Nano through AI Core with custom in-app deployment using LiteRT-LM, MediaPipe, and LiteRT; discusses privacy, latency, offline operation, hardware acceleration, Gemma 4, and compact vision-language models; and closes with audience questions about agent-skill scaling and migration from MediaPipe to LiteRT-LM."
  },
  {
    "slug": "from-agent-traces-to-agent-simulations-rustem-feyzkhanov-snorkel-ai",
    "title": "From Agent Traces to Agent Simulations — Rustem Feyzkhanov, Snorkel AI",
    "url": "https://ai.engineer/talks/from-agent-traces-to-agent-simulations-rustem-feyzkhanov-snorkel-ai",
    "videoId": "Ib5t2RLtxvM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1224000,
    "speakers": [
      {
        "slug": "rustem-feyzkhanov",
        "name": "Rustem Feyzkhanov",
        "organization": "Snorkel AI"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Snorkel AI platform leader Rustem Feyzkhanov explains how production agent traces can become repeatable, company-specific simulation benchmarks that recreate tools, APIs, files, database state, and user interactions. He describes containerized task environments, oracle solutions, final-state verifiers, LLM judges, and long-horizon workflows, then shows how these evaluations support regression gates, model comparison, cost and latency optimization, and an observability-to-experimentation loop. Audience questions address benchmark composition and verifier design."
  },
  {
    "slug": "from-ambient-documentation-to-clinical-intelligence",
    "title": "From Ambient Documentation to Clinical Intelligence",
    "url": "https://ai.engineer/talks/from-ambient-documentation-to-clinical-intelligence",
    "videoId": "u6q-byPWUuo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1295000,
    "speakers": [
      {
        "slug": "chaitanya-asawa",
        "name": "Chaitanya Asawa",
        "organization": "Abridge"
      }
    ],
    "topics": [
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "Abridge engineer Chaitanya Asawa explains how ambient clinical documentation can expand into context-aware clinical intelligence and decision support. Clinician testimonials illustrate reduced documentation burden, while product demonstrations show voice-driven clinical-trial eligibility checks and chart preparation. Asawa traces his background at Vicarious and Glean and argues that healthcare AI requires rigorous evaluation because errors carry serious consequences and clinically valid answers can be almost as difficult to verify as to generate."
  },
  {
    "slug": "from-arc-to-dia-lessons-learned-in-building-ai-browser",
    "title": "From Arc to Dia: Lessons learned in building AI Browser",
    "url": "https://ai.engineer/talks/from-arc-to-dia-lessons-learned-in-building-ai-browser",
    "videoId": "o4scJaQgnFA",
    "event": "AI Engineer Code 2025",
    "durationMs": 1068000,
    "speakers": [
      {
        "slug": "samir-mody",
        "name": "Samir Mody",
        "organization": "The Browser Company"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Samir Mody explains how The Browser Company moved from Arc to Dia, an AI-native browser, and redesigned its engineering workflow around rapid experimentation. He describes embedding prototyping tools directly in the browser, collecting evaluation and training data, enabling nonengineers to test context-aware ideas, and using GEPA to optimize prompts. The talk also highlights AI-browser security risks, including personal-data exfiltration through URL parameters, and confirmation steps for consequential actions."
  },
  {
    "slug": "from-blind-spots-to-merged-prs-continuous-agentic-performance-optimization",
    "title": "From Blind Spots to Merged PRs: Continuous Agentic Performance Optimization",
    "url": "https://ai.engineer/talks/from-blind-spots-to-merged-prs-continuous-agentic-performance-optimization",
    "videoId": "JJGbw4ggaFs",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1366000,
    "speakers": [
      {
        "slug": "may-walter",
        "name": "May Walter",
        "organization": "Hud"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Hud co-founder and CTO May Walter describes a production workflow that continuously discovers and prioritizes performance improvements using function-level runtime intelligence. A weekly GitHub Actions workflow combines GitHub Agentic Workflows, Claude Code, and MCP-accessible production context to analyze repositories, score opportunities by implementation effort and business impact, and deliver actionable reports to Slack. The talk also examines ClickHouse query complexity, unreliable agent-generated fixes, and the disconnect between service-level metrics and code-level diagnosis."
  },
  {
    "slug": "from-chaos-to-choreography-multi-agent-orchestration-patterns-that-actually-work-sandipan-bhaumi",
    "title": "From Chaos to Choreography: Multi-Agent Orchestration Patterns That Actually Work — Sandipan Bhaumik",
    "url": "https://ai.engineer/talks/from-chaos-to-choreography-multi-agent-orchestration-patterns-that-actually-work-sandipan-bhaumi",
    "videoId": "2czYyrTzILg",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1589000,
    "speakers": [
      {
        "slug": "sandipan-bhaumik",
        "name": "Sandipan Bhaumik",
        "organization": "Databricks"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Databricks practitioner Sandipan Bhaumik argues that scaling from one AI agent to many creates a distributed-systems problem rather than merely an LLM problem. Using a stale PostgreSQL-cache incident, he explains race conditions and inconsistent agent decisions, contrasts decentralized event-driven choreography with centralized orchestration, and illustrates orchestration with LangGraph. He recommends immutable versioned agent state and explicit handoffs, then covers circuit breakers, MLflow-based observability, compensating transactions, and rollback for reliable production workflows."
  },
  {
    "slug": "from-copilot-to-colleague-building-trustworthy-productivity-agents-for-high-stakes-work",
    "title": "From Copilot to Colleague: Building Trustworthy Productivity Agents for High-Stakes Work",
    "url": "https://ai.engineer/talks/from-copilot-to-colleague-building-trustworthy-productivity-agents-for-high-stakes-work",
    "videoId": "kDEvo2__Ijg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1185000,
    "speakers": [
      {
        "slug": "joel-hron",
        "name": "Joel Hron",
        "organization": "Thomson Reuters"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Thomson Reuters CTO Joel Hron explains the transition from helpful generative assistants to productive agents operating in high-stakes legal, tax, compliance, audit, and risk workflows. He presents autonomy, context, memory, and coordination as adjustable dimensions governed by user risk tolerance, discusses domain expertise and proprietary knowledge, and describes agent-driven tax preparation, adaptive replanning, and the reuse of legacy applications. The recording concludes with an audience question."
  },
  {
    "slug": "from-fork-to-fleet-designing-an-agent-sandbox-cloud-abhishek-bhardwaj-openai",
    "title": "From fork() to Fleet: Designing an Agent Sandbox Cloud — Abhishek Bhardwaj, OpenAI",
    "url": "https://ai.engineer/talks/from-fork-to-fleet-designing-an-agent-sandbox-cloud-abhishek-bhardwaj-openai",
    "videoId": "OqM67QG_Ikk",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2674000,
    "speakers": [
      {
        "slug": "abhishek-bhardwaj",
        "name": "Abhishek Bhardwaj",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "OpenAI infrastructure engineer Abhishek Bhardwaj explains how to build secure, scalable cloud sandboxes for AI agents and reinforcement-learning workloads. Starting with Linux threads, system calls, privilege boundaries, and container namespaces, he examines isolation and virtual-machine-monitor architecture before arguing that persistent disks and efficient snapshots enable reliable, long-running agent work. He concludes with fleet-scale orchestration and snapshot-assisted sandbox provisioning."
  },
  {
    "slug": "from-hype-to-habit-how-we-re-building-an-ai-first-saas-company-while-still-shipping-the-roadmap",
    "title": "From Hype to Habit: How We’re Building an AI-First SaaS Company—While Still Shipping the Roadmap",
    "url": "https://ai.engineer/talks/from-hype-to-habit-how-we-re-building-an-ai-first-saas-company-while-still-shipping-the-roadmap",
    "videoId": "3YGRcgZJ3yc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1169000,
    "speakers": [
      {
        "slug": "rossella-blatt-vital",
        "name": "Rossella Blatt Vital",
        "organization": "Sprout Social"
      },
      {
        "slug": "deepsha-menghani",
        "name": "Deepsha Menghani",
        "organization": "Sprout Social"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Sprout Social presenters Rossella Blatt Vital and Deepsha Menghani describe moving an established SaaS business beyond isolated AI features toward an AI-first operating model while continuing to deliver its roadmap. They outline a multidimensional transformation framework, use an illustrative puppy-health application to explain incremental customer value, and discuss learning-focused MVPs, organizational process tradeoffs, company-wide AI fluency, and T-shaped technical talent."
  },
  {
    "slug": "from-mcp-to-scale-pipelines-that-build-themselves",
    "title": "From MCP to Scale: Pipelines That Build Themselves",
    "url": "https://ai.engineer/talks/from-mcp-to-scale-pipelines-that-build-themselves",
    "videoId": "zTZ0qunQXnM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1526000,
    "speakers": [
      {
        "slug": "rafael-levi",
        "name": "Rafael Levi",
        "organization": "Bright Data"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Rafael Levi demonstrates how Bright Data MCP, Claude Code, and the brightdata/skills repository enable agents to inspect websites, extract HTML and selectors, generate reusable scrapers, and maintain self-healing collection pipelines. The talk contrasts repeated LLM-based page parsing with token-efficient scripts, discusses anti-bot and remote-browser infrastructure, and describes scheduled monitoring and automated scraper maintenance."
  },
  {
    "slug": "from-mixture-of-experts-to-mixture-of-agents-with-super-fast-inference",
    "title": "From Mixture of Experts to Mixture of Agents … with Super Fast Inference",
    "url": "https://ai.engineer/talks/from-mixture-of-experts-to-mixture-of-agents-with-super-fast-inference",
    "videoId": "tzRvcTEapzo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 3195000,
    "speakers": [
      {
        "slug": "daniel-kim",
        "name": "Daniel Kim",
        "organization": "Cerebras"
      },
      {
        "slug": "daria-soboleva",
        "name": "Daria Soboleva",
        "organization": "Cerebras"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Cerebras presenters Daniel Kim and Daria Soboleva explain how Mixture of Experts supports efficient language-model scaling and introduce Mixture of Agents as a collaborative multi-model inference architecture. The hands-on workshop covers obtaining a Cerebras API key, deploying a GitHub-based Streamlit application, configuring prompts, and participating in an interactive coding competition, followed by audience questions about model architectures, fine-tuning, and real-time or multimodal APIs."
  },
  {
    "slug": "from-model-weights-to-api-endpoint-with-tensorrt-llm",
    "title": "From model weights to API endpoint with TensorRT-LLM",
    "url": "https://ai.engineer/talks/from-model-weights-to-api-endpoint-with-tensorrt-llm",
    "videoId": "Lko9lTGD_9U",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 6001000,
    "speakers": [
      {
        "slug": "philip-kiely",
        "name": "Philip Kiely",
        "organization": "Baseten"
      },
      {
        "slug": "pankaj-gupta",
        "name": "Pankaj Gupta",
        "organization": "Baseten"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Baseten presenters Philip Kiely and Pankaj Gupta lead a hands-on workshop on taking language-model weights to a production API with TensorRT-LLM. They cover supported-model selection, cloud-GPU setup, engine building and automated configuration, FP8 quantization of weights and the KV cache, deployment with Baseten and Truss, and benchmarking metrics including time to first token. Audience questions address alternatives such as vLLM and practical deployment concerns."
  },
  {
    "slug": "from-pm-at-stripe-to-building-an-ai-startup-a-recent-founder-s-journey-mounir-mouawad",
    "title": "From PM at Stripe to Building an AI Startup, a Recent Founder's Journey - Mounir Mouawad",
    "url": "https://ai.engineer/talks/from-pm-at-stripe-to-building-an-ai-startup-a-recent-founder-s-journey-mounir-mouawad",
    "videoId": "HIGpxVjGFBw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 719000,
    "speakers": [
      {
        "slug": "mounir-mouawad",
        "name": "Mounir Mouawad",
        "organization": "Portia AI"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "Portia AI co-founder Mounir Mouawad reflects on leaving product roles at Stripe, Google, and Amazon to build an open-source SDK for production AI agents in regulated industries. Using video-game analogies, he contrasts established-company product development with the uncertainty of discovering emerging AI customer needs, the exhilaration and necessity of rapid iteration around standards such as MCP, and the challenge of earning distribution through partnerships, integration documentation, and developer communities without an established brand."
  },
  {
    "slug": "from-rl-to-irl-gaurav-mishra-amazon-agi-lab",
    "title": "From RL to IRL — Gaurav Mishra, Amazon AGI Lab",
    "url": "https://ai.engineer/talks/from-rl-to-irl-gaurav-mishra-amazon-agi-lab",
    "videoId": "Cc0_nyxROBA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1066000,
    "speakers": [
      {
        "slug": "gaurav-mishra",
        "name": "Gaurav Mishra",
        "organization": "Amazon AGI Lab"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Amazon AGI Lab researcher Gaurav Mishra explains why reinforcement-learning successes with coding agents do not directly translate into reliable real-world computer use. He contrasts RL with supervised fine-tuning, describes browser-agent interfaces including MCP and Playwright, and examines partial observability, misleading web content, and ambiguous task completion. His proposed safeguards include realistic digital sandboxes, recovery-aware training, process reward models, and action-risk classification."
  },
  {
    "slug": "from-self-driving-to-autonomous-voice-agents-brooke-hopkins-coval",
    "title": "From Self-driving to Autonomous Voice Agents — Brooke Hopkins, Coval",
    "url": "https://ai.engineer/talks/from-self-driving-to-autonomous-voice-agents-brooke-hopkins-coval",
    "videoId": "kDczF4wBh8s",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1172000,
    "speakers": [
      {
        "slug": "brooke-hopkins",
        "name": "Brooke Hopkins",
        "organization": "Coval"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Coval founder Brooke Hopkins explains how her experience building simulation and evaluation infrastructure at Waymo informs reliable deployment of autonomous voice agents. She argues that large-scale, probabilistic simulations can reconcile agent autonomy with production reliability, and outlines scenario coverage, continuous CI/CD evaluation, product-specific success metrics, public benchmarks, and carefully calibrated LLM-as-a-judge workflows."
  },
  {
    "slug": "from-signal-to-pr-anatomy-of-a-self-improving-agent",
    "title": "From Signal to PR: Anatomy of a Self-Improving Agent",
    "url": "https://ai.engineer/talks/from-signal-to-pr-anatomy-of-a-self-improving-agent",
    "videoId": "9HbzAWnKbo4",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1236000,
    "speakers": [
      {
        "slug": "jason-lopatecki",
        "name": "Jason Lopatecki",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Arize co-founder Jason Lopatecki describes Signal, an Arize AX capability that combines production traces, observability skills, automated triggers, and repository access to investigate AI-agent failures and help prepare fixes for human review. Using Arize's Alyx assistant as a debugging example, he contrasts AX with open-source Phoenix, discusses enterprise VPC deployment, and answers audience questions about Claude Code and evaluations."
  },
  {
    "slug": "from-software-developer-to-ai-engineer",
    "title": "From Software Developer to AI Engineer",
    "url": "https://ai.engineer/talks/from-software-developer-to-ai-engineer",
    "videoId": "90Ir4TMMbS0",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1188000,
    "speakers": [
      {
        "slug": "antje-barth",
        "name": "Antje Barth",
        "organization": "AWS"
      },
      {
        "slug": "mike-chambers",
        "name": "Mike Chambers",
        "organization": "AWS"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "AWS developer advocate Antje Barth outlines practical steps for software developers becoming AI engineers: learn generative-AI fundamentals, adopt AI development assistants, and prototype applications. A demonstration introduced as Mike Chambers showcases Amazon Q Developer for IDE assistance, serverless development with AWS SAM CLI, code explanation and generation, and DynamoDB feature implementation. The keynote also discusses fine-tuning, RAG, agent action groups illustrated with Minecraft, and local AI-engineering community initiatives."
  },
  {
    "slug": "from-stateless-nightmares-to-durable-agents",
    "title": "From Stateless Nightmares to Durable Agents",
    "url": "https://ai.engineer/talks/from-stateless-nightmares-to-durable-agents",
    "videoId": "flf_IKnFYnE",
    "event": "AI Engineer Code 2025",
    "durationMs": 1333000,
    "speakers": [
      {
        "slug": "samuel-colvin",
        "name": "Samuel Colvin",
        "organization": "Pydantic"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Samuel Colvin demonstrates how Pydantic AI and Temporal make long-running, multi-agent workflows resilient to crashes and restarts. Using a two-agent twenty-questions example and a deep-research workflow, he explains Temporal activity wrapping, persisted workflow state, workflow-ID-based resumption, and replay of previously completed LLM calls. He also discusses Pydantic Logfire, Pydantic Evals, model comparisons, execution costs, and a reported Vertex SDK integration limitation."
  },
  {
    "slug": "from-systems-of-record-to-systems-of-context",
    "title": "From Systems of Record to Systems of Context",
    "url": "https://ai.engineer/talks/from-systems-of-record-to-systems-of-context",
    "videoId": "Btk8wDUVs74",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 958000,
    "speakers": [
      {
        "slug": "omri-bruchim",
        "name": "Omri Bruchim",
        "organization": "Monday"
      },
      {
        "slug": "tomer-ast",
        "name": "Tomer Ast",
        "organization": "monday.com"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "monday.com engineering managers Omri Bruchim and Tomer Ast explain why enterprise AI assistants cannot prioritize work merely by retrieving disconnected tasks, messages, meetings, and documents. Their proposed Monday world model precomputes relationships and user context for monday sidekick, combining durable offline behavioral patterns with current commitments and resilient serving that isolates faulty sources, verifies live data, and falls back to previously verified context."
  },
  {
    "slug": "from-text-to-vision-to-voice-exploring-multimodality-with-openai",
    "title": "From Text to Vision to Voice: Exploring Multimodality with OpenAI",
    "url": "https://ai.engineer/talks/from-text-to-vision-to-voice-exploring-multimodality-with-openai",
    "videoId": "yJHw33cVeHo",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1419000,
    "speakers": [
      {
        "slug": "romain-huet",
        "name": "Romain Huet",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "OpenAI's Romain Huet traces the developer platform's progression from GPT-3 and text applications through GPT-4 vision and GPT-4 Turbo to GPT-4o, an omni model combining real-time text, audio, and visual understanding. Live conversational and visual demonstrations illustrate multimodal interaction, followed by discussion of cheaper and customizable models, agents, and opportunities for developers building AI-native products."
  },
  {
    "slug": "from-tokens-to-cells-foundation-models-for-single-cell-biology-akram-baharlouei-altos-labs",
    "title": "From Tokens to Cells: Foundation Models for Single-Cell Biology - Akram Baharlouei, Altos Labs",
    "url": "https://ai.engineer/talks/from-tokens-to-cells-foundation-models-for-single-cell-biology-akram-baharlouei-altos-labs",
    "videoId": "-561cZmir5Q",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1017000,
    "speakers": [
      {
        "slug": "akram-baharlouei",
        "name": "Akram Baharlouei",
        "organization": "Altos Labs"
      }
    ],
    "topics": [
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Altos Labs machine learning engineer Akram Baharlouei explains the engineering challenges of applying foundation models to single-cell biology, connecting cellular rejuvenation and reprogramming to drug development and virtual-cell modeling. She examines RNA-seq scale, measurement limitations, and multimodal data, then discusses transformer-based models including scGPT and Geneformer, benchmarking with scGeneScope and PerturBench, and a model described in the talk as Prime Flow."
  },
  {
    "slug": "from-transcription-to-live-music-gemini-s-audio-stack-thor-schaeff-google-deepmind",
    "title": "From Transcription to Live Music: Gemini's Audio Stack — Thor Schaeff, Google DeepMind",
    "url": "https://ai.engineer/talks/from-transcription-to-live-music-gemini-s-audio-stack-thor-schaeff-google-deepmind",
    "videoId": "Bc6Ojl2XS1w",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1174000,
    "speakers": [
      {
        "slug": "thor-schaeff",
        "name": "Thor Schaeff",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Google DeepMind developer-relations engineer Thor Schaeff demonstrates Gemini's audio stack: Gemma 4 and Veo 3.1 Lite; EchoScript's single-call audio analysis with language, emotion, translation, summaries, and structured outputs; prompt-directed synthetic speech; Gemini 3.1 Flash Live's real-time multimodal conversations; and an interactive music-generation finale."
  },
  {
    "slug": "from-vibe-coding-to-vibe-engineering",
    "title": "From Vibe Coding to Vibe Engineering",
    "url": "https://ai.engineer/talks/from-vibe-coding-to-vibe-engineering",
    "videoId": "JV-wY5pxXLo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1528000,
    "speakers": [
      {
        "slug": "kitze",
        "name": "Kitze",
        "organization": "Sizzy"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Kitze, founder of developer-browser product Sizzy, uses frontend-development history and industry satire to distinguish casual vibe coding from engineering effectively with large language models. He discusses practical judgment about generated code, production-data risks, rapidly changing model choices, Cursor and Composer 1, MCP hype, and how AI agents may reshape developer work."
  },
  {
    "slug": "from-writing-code-to-designing-systems-how-the-developer-role-is-changing-chris-noring-microsoft",
    "title": "From Writing Code to Designing Systems: How the Developer Role is Changing — Chris Noring, Microsoft",
    "url": "https://ai.engineer/talks/from-writing-code-to-designing-systems-how-the-developer-role-is-changing-chris-noring-microsoft",
    "videoId": "GdvKNwMcfd0",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1385000,
    "speakers": [
      {
        "slug": "chris-noring",
        "name": "Chris Noring",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Microsoft developer advocate Chris Noring argues that AI-assisted development shifts engineers from writing every line of code toward designing systems, planning work, and coordinating coding agents. He outlines a CLI-first workflow spanning GitHub Copilot CLI, editors, and GitHub interfaces, contrasts Copilot with Claude-based tools, and describes scaling work across issues and pull requests. Agent Skills, SKILL.md instructions, AGENTS.md, custom agents, and architectural guardrails help preserve consistency and prevent increased output from becoming low-quality generated code."
  },
  {
    "slug": "frontier-ai-at-home-literally",
    "title": "Frontier AI at Home (literally)",
    "url": "https://ai.engineer/talks/frontier-ai-at-home-literally",
    "videoId": "ESbWpPT_9-o",
    "event": "AI Engineer Europe 2026",
    "durationMs": 6302000,
    "speakers": [
      {
        "slug": "alex-cheema",
        "name": "Alex Cheema",
        "organization": "EXO Labs"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "EXO Labs co-founder Alex Cheema presents a hands-on workshop about running frontier language models on local hardware. He introduces prefill and decode, argues for local ownership and privacy, discusses energy-efficient hardware-software co-design and Mac-based infrastructure, and explores open-model harnesses, test-time compute, quantization tradeoffs, cluster dashboards, and large-prompt demonstrations through extensive audience discussion."
  },
  {
    "slug": "frontier-feud",
    "title": "Frontier Feud",
    "url": "https://ai.engineer/talks/frontier-feud",
    "videoId": "BWqB6aatreE",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1346000,
    "speakers": [
      {
        "slug": "barr-yaron",
        "name": "Barr Yaron",
        "organization": "Amplify Partners"
      },
      {
        "slug": "mihir",
        "name": "Mihir",
        "organization": "Anthropic"
      },
      {
        "slug": "john",
        "name": "John",
        "organization": "Anthropic"
      },
      {
        "slug": "tina",
        "name": "Tina",
        "organization": "Reflection AI"
      },
      {
        "slug": "shresta",
        "name": "Shresta"
      },
      {
        "slug": "paige",
        "name": "Paige",
        "organization": "Google DeepMind"
      },
      {
        "slug": "colin",
        "name": "Colin",
        "organization": "Augment Code"
      },
      {
        "slug": "petra",
        "name": "Petra",
        "organization": "Google"
      },
      {
        "slug": "steven",
        "name": "Steven",
        "organization": "Thinking Machines"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Amplify Partners' Barr Yaron hosts a Family Feud-style contest based on answers from 100 AI engineers. Participants identifying affiliations with Anthropic, Reflection AI, Google DeepMind, Augment Code, Google, and Thinking Machines introduce themselves and share predictions about frontier-model training, coding agents, on-device deployment, factuality, and future model architectures. Teams compete over influential AI researchers and technical references before a Fast Money round; Yaron closes by inviting participation in a state-of-AI-engineering survey."
  },
  {
    "slug": "frontier-results-on-device-rl-nabors-arize",
    "title": "Frontier results, on device - RL Nabors, Arize",
    "url": "https://ai.engineer/talks/frontier-results-on-device-rl-nabors-arize",
    "videoId": "fWXJM-J0ZB8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1852000,
    "speakers": [
      {
        "slug": "rachel-lee-nabors-rl-nabors",
        "name": "Rachel Lee Nabors (RL Nabors)",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Rachel Lee Nabors of Arize explains how teams can replace selected cloud frontier-model calls with task-specific and on-device models to reduce privacy exposure, latency, and inference costs. Using Mima as a practical example, the talk compares Qwen, Llama, and Gemma candidates, iterates on prompt structure, and uses Phoenix-supported evaluation metrics to assess output quality and performance."
  },
  {
    "slug": "full-spec-mcp-hidden-capabilities-of-the-mcp-spec-harald-kirschner-microsoft-vs-code",
    "title": "Full Spec MCP: Hidden Capabilities of the MCP Spec — Harald Kirschner, Microsoft/VS Code",
    "url": "https://ai.engineer/talks/full-spec-mcp-hidden-capabilities-of-the-mcp-spec-harald-kirschner-microsoft-vs-code",
    "videoId": "ExeD-8gFUMM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 893000,
    "speakers": [
      {
        "slug": "harald-kirschner",
        "name": "Harald Kirschner",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Harald Kirschner explains why treating Model Context Protocol as merely an API-and-tools wrapper misses capabilities available through fuller client-server implementations. Using VS Code examples, he discusses dynamic discovery and workspace roots, tool overload and per-chat selection, framework-aware context, resource summarization through sampling, improved authorization, and forthcoming user elicitation, while encouraging developers to test draft protocol features and provide feedback."
  },
  {
    "slug": "full-workshop-agent-auth-protocol-paola-estefania-de-campos-better-auth",
    "title": "Full Workshop: Agent Auth Protocol — Paola Estefanía de Campos, Better Auth",
    "url": "https://ai.engineer/talks/full-workshop-agent-auth-protocol-paola-estefania-de-campos-better-auth",
    "videoId": "JvKO40CFq-s",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2456000,
    "speakers": [
      {
        "slug": "paola-estefania-de-campos",
        "name": "Paola Estefanía de Campos",
        "organization": "Better Auth"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Paola Estefanía de Campos presents Better Auth’s Agent Auth Protocol, explaining why AI agents need distinct identities and narrowly scoped capabilities instead of shared user credentials or unrestricted tokens. The interactive workshop covers service discovery, endpoint-to-capability mapping, approvals, agent hosts, revocation, auditability, and lifecycle traceability, with audience questions about enterprise deployment, APIs, SDKs, and related standards. Anticipated co-presenter Bereket was absent."
  },
  {
    "slug": "full-workshop-realtime-voice-ai-mark-backman-daily",
    "title": "Full Workshop: Realtime Voice AI — Mark Backman, Daily",
    "url": "https://ai.engineer/talks/full-workshop-realtime-voice-ai-mark-backman-daily",
    "videoId": "nxuTVd7v7dg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 4181000,
    "speakers": [
      {
        "slug": "mark-backman",
        "name": "Mark Backman",
        "organization": "Daily"
      },
      {
        "slug": "aleix",
        "name": "Aleix",
        "organization": "Daily"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Mark Backman and an additional presenter introduced as Aleix lead a hands-on Daily workshop on building realtime voice agents with the open-source Pipecat framework and Gemini Live. They introduce the public workshop repository, Python setup, Google and Daily API credentials, WebRTC/WebSocket transport, composable audio-processing pipelines, interchangeable speech and language models, voice activity detection, semantic turn-taking, and a live conversational demonstration."
  },
  {
    "slug": "full-workshop-setting-yourself-up-for-success-jason-liu-openai-codex",
    "title": "Full Workshop: Setting Yourself Up for Success — Jason Liu, OpenAI Codex",
    "url": "https://ai.engineer/talks/full-workshop-setting-yourself-up-for-success-jason-liu-openai-codex",
    "videoId": "il1c1a2FufU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 4502000,
    "speakers": [
      {
        "slug": "jason-liu",
        "name": "Jason Liu",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "OpenAI developer-experience engineer Jason Liu leads an interactive workshop on using Codex as a general-purpose assistant for computer use and sustained knowledge work. He demonstrates persistent memory systems, context compaction, parallel threads and sub-agents, Slack-based automation, personal knowledge and relationship tracking, long-running task logs, mobile-assisted computer control, and practical reasoning-effort choices, with audience questions throughout."
  },
  {
    "slug": "fun-stories-from-building-openrouter-and-where-all-this-is-going",
    "title": "fun stories from building OpenRouter and where all this is going",
    "url": "https://ai.engineer/talks/fun-stories-from-building-openrouter-and-where-all-this-is-going",
    "videoId": "84Vtz2IL1Ug",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1127000,
    "speakers": [
      {
        "slug": "alex-atallah",
        "name": "Alex Atallah",
        "organization": "OpenRouter"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "OpenRouter co-founder Alex Atallah recounts how the emergence of open-weight models and Stanford Alpaca challenged assumptions that AI inference would become a winner-take-all market. He traces the path from the Window AI browser extension to OpenRouter’s multi-provider model marketplace, then discusses provider fallback, inference plugins, MCP integration, improved model discovery, prompt observability, and reducing vendor lock-in."
  },
  {
    "slug": "function-calling-is-all-you-need",
    "title": "Function Calling is All You Need",
    "url": "https://ai.engineer/talks/function-calling-is-all-you-need",
    "videoId": "KUEmEb71vzQ",
    "event": "AI Engineer Summit 2025",
    "durationMs": 6174000,
    "speakers": [
      {
        "slug": "ilan-bigio",
        "name": "Ilan Bigio",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "OpenAI developer experience engineer Ilan Bigio leads a remotely delivered, interactive coding workshop arguing that function calling underpins practical AI agents. He traces the progression from GPT base models through InstructGPT and WebGPT, implements a tool-calling agent loop, discusses retrieval, memory and application state, explores asynchronous execution and agent frameworks including Swarm and Pydantic AI, and demonstrates agents that write their own tools."
  },
  {
    "slug": "future-proof-coding-agents",
    "title": "Future-Proof Coding Agents",
    "url": "https://ai.engineer/talks/future-proof-coding-agents",
    "videoId": "wVl6ZjELpBk",
    "event": "AI Engineer Code 2025",
    "durationMs": 1068000,
    "speakers": [
      {
        "slug": "bill-chen",
        "name": "Bill Chen",
        "organization": "OpenAI"
      },
      {
        "slug": "brian-fioca",
        "name": "Brian Fioca",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "OpenAI presenters Bill Chen and Brian Fioca explain how to build durable coding agents around three components—interface, model, and harness—and why rapidly changing models, APIs, tools, prompting conventions, and context windows make custom harnesses difficult to maintain. They present Codex and GPT-5.1-Codex-Max as reusable agent infrastructure with automatic context compaction, IDE and command-line access, cloud execution, and integrations including ChatGPT, Slack, and GitHub, allowing product teams to focus on differentiated workflows and applications beyond coding."
  },
  {
    "slug": "fuzzing-in-the-genai-era",
    "title": "Fuzzing in the GenAI Era",
    "url": "https://ai.engineer/talks/fuzzing-in-the-genai-era",
    "videoId": "OMGPvW8TBHc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1152000,
    "speakers": [
      {
        "slug": "leonard-tang",
        "name": "Leonard Tang",
        "organization": "Haize Labs"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Leonard Tang of Haize Labs argues that static golden-dataset evaluations fail to capture the brittleness preventing generative-AI applications from reaching production. He presents Haizing, an AI fuzzing and property-based testing approach combining diverse simulated inputs with automated quality judgments, discusses Verdict's composable LLM-as-a-judge architecture and reinforcement-learning-trained j1-micro reward model, and describes evaluating enterprise banking voice agents across text, audio, and multi-turn conversations."
  },
  {
    "slug": "gadgets-personal-app-vibe-coding-that-is-actually-safe",
    "title": "Gadgets: Personal app vibe coding that is actually safe",
    "url": "https://ai.engineer/talks/gadgets-personal-app-vibe-coding-that-is-actually-safe",
    "videoId": "RmS5s6Wbin4",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1134000,
    "speakers": [
      {
        "slug": "kenton-varda",
        "name": "Kenton Varda",
        "organization": "Cloudflare"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Cloudflare's Kenton Varda argues that personal AI-generated applications require a different infrastructure model from conventional centrally hosted software. He demonstrates Gadgets for collaborative whiteboarding, email filtering, and GitHub pull-request review, then describes an isolation architecture using null-origin sandboxed iframes, Content Security Policy, postMessage, Cap'n Web RPC, Durable Objects, and Cloudflare Workers to constrain potentially unsafe AI-generated client and server code."
  },
  {
    "slug": "gateways-are-all-you-need",
    "title": "Gateways are All You Need",
    "url": "https://ai.engineer/talks/gateways-are-all-you-need",
    "videoId": "CD6R4Wf3jnY",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1068000,
    "speakers": [
      {
        "slug": "karan-sampath",
        "name": "Karan Sampath",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Anthropic product engineer Karan Sampath argues that enterprise adoption of the Model Context Protocol is constrained by weak observability, fragmented access control, security concerns, and deployment bottlenecks. He proposes shared MCP gateways that centralize authentication, role-based permissions, developer tooling, and organizational visibility, establishing a root of trust while enabling teams to iterate independently. The longer-term architectural goal is to separate agent harnesses from enterprise data layers."
  },
  {
    "slug": "gemini-nano-on-device-florina-muntenescu-oli-gaymond-google-deepmind",
    "title": "Gemini Nano on device — Florina Muntenescu & Oli Gaymond, Google DeepMind",
    "url": "https://ai.engineer/talks/gemini-nano-on-device-florina-muntenescu-oli-gaymond-google-deepmind",
    "videoId": "owH1f0N-keY",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1178000,
    "speakers": [
      {
        "slug": "florina-muntenescu",
        "name": "Florina Muntenescu",
        "organization": "Google DeepMind"
      },
      {
        "slug": "oli-gaymond",
        "name": "Oli Gaymond",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Florina Muntenescu and Oli Gaymond introduce Android's on-device, hybrid, and cloud AI options, focusing on Gemini Nano through ML Kit GenAI APIs and the shared AICore system service. They contrast that managed approach with customizable LiteRT models and take audience questions about memory, battery consumption, device compatibility, request queuing, and AI Edge Gallery."
  },
  {
    "slug": "gemma-4-deep-dive-cassidy-hardin-google-deepmind",
    "title": "Gemma 4 Deep Dive — Cassidy Hardin, Google DeepMind",
    "url": "https://ai.engineer/talks/gemma-4-deep-dive-cassidy-hardin-google-deepmind",
    "videoId": "_A367W_qvc8",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1143000,
    "speakers": [
      {
        "slug": "cassidy-hardin",
        "name": "Cassidy Hardin",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Google DeepMind researcher Cassidy Hardin explains Gemma 4’s original four-model lineup, including E2B and E4B on-device models, a 26B mixture-of-experts model, and a 31B dense model, alongside its Apache 2.0 licensing. She details interleaved local-global and grouped-query attention, flash-resident Per Layer Embeddings that reduce on-device VRAM pressure, variable-resolution vision processing, conformer-based audio encoding, and agentic deployment through Google AI Studio and Vertex AI."
  },
  {
    "slug": "gemma-deepmind-s-family-of-open-models",
    "title": "Gemma, DeepMind's Family of Open Models",
    "url": "https://ai.engineer/talks/gemma-deepmind-s-family-of-open-models",
    "videoId": "_gVFUEdhCyI",
    "event": "AI Engineer Europe 2026",
    "durationMs": 926000,
    "speakers": [
      {
        "slug": "omar-sanseviero",
        "name": "Omar Sanseviero",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "A Google DeepMind keynote introduces Gemma 4 and contrasts it with Gemma 3, emphasizing open, multimodal models that run on phones, laptops, Raspberry Pi devices, and consumer GPUs. Demonstrations and technical discussion cover mixture-of-experts and 31B variants, parallel offline agents, LM Arena comparisons, and E2B/E4B per-layer embeddings that can be offloaded with llama.cpp. The talk also surveys integrations across the open-source ecosystem, offline Android development, ShieldGemma, and community adaptation for Southeast Asian languages."
  },
  {
    "slug": "general-purpose-robots-as-professional-chefs",
    "title": "General purpose robots as professional Chefs",
    "url": "https://ai.engineer/talks/general-purpose-robots-as-professional-chefs",
    "videoId": "MBWGiWJDlSo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1138000,
    "speakers": [
      {
        "slug": "nikhil-abraham",
        "name": "Nikhil Abraham",
        "organization": "CloudChef"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "CloudChef co-founder and CEO Nikhil Abraham explains how the company turns general-purpose, two-armed mobile robots into commercial kitchen workers that learn recipes from a single demonstration and adapt to unfamiliar kitchens. He discusses culinary perception and thermal sensing, comparisons with Gemini 2.5 and o3, manipulation evaluations, a reported 95% autonomous and 5% teleoperated operating mix, hourly deployment economics, and real customer installations. An extended audience Q&A covers blind taste-test benchmarking, robot hardware, appliance compatibility, operating speed, restaurant availability, contamination risks, and ingredient replenishment."
  },
  {
    "slug": "generative-video-at-the-speed-of-light",
    "title": "Generative Video at the Speed of Light",
    "url": "https://ai.engineer/talks/generative-video-at-the-speed-of-light",
    "videoId": "Xln-On3syJk",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 523000,
    "speakers": [
      {
        "slug": "keegan-mccallum",
        "name": "Keegan McCallum",
        "organization": "uRun"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "uRun founder Keegan McCallum discusses how generative video is advancing in efficiency and continuous, real-time output alongside visual quality. Using Helios and its Wan 2.1 14B lineage, he explores long-horizon generation, declining usage costs, interactive avatars, accessibility, distributed GPU deployment, WebRTC networking, and agent-oriented application development through a CLI or MCP server."
  },
  {
    "slug": "git-push-get-an-ai-api",
    "title": "Git push, get an AI API.",
    "url": "https://ai.engineer/talks/git-push-get-an-ai-api",
    "videoId": "R7xlYs5p0yA",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 2701000,
    "speakers": [
      {
        "slug": "ryan-fox-tyler",
        "name": "Ryan Fox-Tyler",
        "organization": "Hypermode"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Ryan Fox-Tyler’s Hypermode workshop demonstrates how to incrementally build and deploy AI-powered application features. Presenters introduce a multiplayer categorization game, then use a GitHub Issues triage template to explore application building blocks, TypeScript development, retrieval-augmented summarization, and platform-hosted issue classification with a compact pretrained Hugging Face model, avoiding the need for participants to provide external API credentials during that exercise."
  },
  {
    "slug": "github-copilot-the-world-s-most-widely-adopted-ai-developer-tool",
    "title": "GitHub Copilot - The World’s Most Widely Adopted AI Developer Tool",
    "url": "https://ai.engineer/talks/github-copilot-the-world-s-most-widely-adopted-ai-developer-tool",
    "videoId": "uhoPM-ABuV0",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1789000,
    "speakers": [
      {
        "slug": "dave-burnison",
        "name": "Dave Burnison",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "GitHub developer advocate Dave Burnison demonstrates how GitHub Copilot expanded from IDE code completion into conversational assistance, enterprise knowledge retrieval, and GitHub.com pull-request workflows. Examples cover supported editors and repository hosts, context-aware Django/Python code generation, Copilot Chat, GitHub Actions and Octopus Deploy integration, and enterprise-managed user isolation."
  },
  {
    "slug": "github-copilot-the-world-s-most-widely-adopted-ai-developer-tool-repeat",
    "title": "GitHub Copilot - The World’s Most Widely Adopted AI Developer Tool (Repeat)",
    "url": "https://ai.engineer/talks/github-copilot-the-world-s-most-widely-adopted-ai-developer-tool-repeat",
    "videoId": "sKiTUuEV1rw",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 4785000,
    "speakers": [
      {
        "slug": "dave-burnison",
        "name": "Dave Burnison",
        "organization": "GitHub"
      },
      {
        "slug": "alex-malebranche",
        "name": "Alex Malebranche",
        "organization": "GitHub"
      },
      {
        "slug": "dimitrios-philliou",
        "name": "Dimitrios Philliou",
        "organization": "GitHub"
      },
      {
        "slug": "christina-warren",
        "name": "Christina Warren",
        "organization": "GitHub"
      },
      {
        "slug": "harald",
        "name": "Harald",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Christina Warren and GitHub colleagues lead a repeat hands-on GitHub Copilot workshop covering repository setup, browser-based GitHub Codespaces, contextual IDE suggestions, Copilot Chat, and workspace-aware chat participants. Demonstrations address precise prompting, inline ghost-text completions, extensible chat participants and models, enterprise knowledge bases, and managing issues and pull requests from Codespaces. The recorded introductions identify an assisting presenter, Harald, who is absent from the authoritative scheduled roster, while scheduled presenter Dimitrios Philliou is not identified in the supplied transcript excerpts."
  },
  {
    "slug": "github-next-explorations",
    "title": "GitHub Next Explorations",
    "url": "https://ai.engineer/talks/github-next-explorations",
    "videoId": "1oySeF37SZc",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1098000,
    "speakers": [
      {
        "slug": "rahul-pandita",
        "name": "Rahul Pandita",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "GitHub researcher Rahul Pandita introduces GitHub Next's approach to exploring future software-engineering tools through prototypes, internal dogfooding, and product incubation. He describes Copilot Next Edit Suggestions, which predicts useful code changes beyond the cursor, and Copilot Workspace, which helps developers move from natural-language specifications through planning and implementation. The presentation emphasizes developer trust, human control, iteration, and collaboration with AI as a thought partner."
  },
  {
    "slug": "github-s-ai-powered-security-platform",
    "title": "GitHub's AI-Powered Security Platform",
    "url": "https://ai.engineer/talks/github-s-ai-powered-security-platform",
    "videoId": "utTqdQpe39A",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1425000,
    "speakers": [
      {
        "slug": "sarah-khalife",
        "name": "Sarah Khalife",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "GitHub principal solutions engineer Sarah Khalife explains how GitHub Advanced Security incorporates AI into application-security workflows, including code scanning, data-flow-based vulnerability detection, secret scanning, and faster remediation. She frames security as a shared developer responsibility and demonstrates repository-level security capabilities alongside the wider GitHub and Copilot platform."
  },
  {
    "slug": "give-your-chat-agent-a-voice",
    "title": "Give Your Chat Agent a Voice",
    "url": "https://ai.engineer/talks/give-your-chat-agent-a-voice",
    "videoId": "DCZZ3AJKzuc",
    "event": "AI Engineer Europe 2026",
    "durationMs": 492000,
    "speakers": [
      {
        "slug": "luke-harries",
        "name": "Luke Harries",
        "organization": "ElevenLabs"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "ElevenLabs Head of Growth Luke Harries previews Voice Engine, a wrapper that adds speech interaction to existing chat agents without replacing their orchestration, retrieval, or tools. He describes Scribe speech recognition, Eleven v3 speech synthesis, emotion-aware turn-taking and semantic VAD, then demonstrates server and client SDKs, shadcn/ui-style components, telephony integrations, and client-side versus server-side tool calling."
  },
  {
    "slug": "giving-a-voice-to-ai-agents",
    "title": "Giving a Voice to AI Agents",
    "url": "https://ai.engineer/talks/giving-a-voice-to-ai-agents",
    "videoId": "heYmh_lsX5s",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 788000,
    "speakers": [
      {
        "slug": "scott-stephenson",
        "name": "Scott Stephenson",
        "organization": "Deepgram"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Deepgram CEO Scott Stephenson explains how modern voice agents combine speech-to-text, large language models, and text-to-speech into responsive conversational systems. He contrasts earlier assistants with open-ended voice AI, highlights a Daily implementation using Deepgram and Llama that approaches 500-millisecond response times, and discusses multimodal context, enterprise controllability, modular architectures, and operating costs."
  },
  {
    "slug": "going-beyond-rag-extended-mind-transformers",
    "title": "Going beyond RAG: Extended Mind Transformers",
    "url": "https://ai.engineer/talks/going-beyond-rag-extended-mind-transformers",
    "videoId": "cS6M6Jec0lU",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 964000,
    "speakers": [
      {
        "slug": "phoebe-klett",
        "name": "Phoebe Klett",
        "organization": "Normal Computing"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Normal Computing researcher Phoebe Klett presents Extended Mind Transformers, which augment decoder-only language models with cached external key-value memories retrieved through model-native query/key similarity, avoiding fine-tuning. She contrasts the approach with conventional RAG, explains RoPE and ALiBi compatibility, presents a counterfactual long-context retrieval benchmark and a reported GPT-4 comparison, and discusses causal citations, active-learning-inspired hallucination reduction, memory-stride tuning, and openly available models and code."
  },
  {
    "slug": "good-design-hasn-t-changed-with-ai",
    "title": "Good design hasn’t changed with AI",
    "url": "https://ai.engineer/talks/good-design-hasn-t-changed-with-ai",
    "videoId": "7e7eVtcygCM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1225000,
    "speakers": [
      {
        "slug": "john-pham",
        "name": "John Pham",
        "organization": "SF Compute"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "John Pham argues that AI changes the pace of software development but not the human-centered fundamentals of good design. Using SF Compute’s GPU-cloud onboarding as a case study, he explains how fast page loads, smooth interactions, small JavaScript bundles, and Core Web Vitals metrics improve perceived quality. He also shows how transparent compliance requirements, clear onboarding steps, local time zones, reduced cognitive load, and GPU-themed visual details make demanding workflows feel trustworthy and delightful."
  },
  {
    "slug": "government-agents-ai-agents-meet-tough-regulations-mark-myshatyn-los-alamos-national-laboratory",
    "title": "Government Agents: AI Agents Meet Tough Regulations — Mark Myshatyn, Los Alamos National Laboratory",
    "url": "https://ai.engineer/talks/government-agents-ai-agents-meet-tough-regulations-mark-myshatyn-los-alamos-national-laboratory",
    "videoId": "TnSGx36Ly0Q",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 991000,
    "speakers": [
      {
        "slug": "mark-myshatyn",
        "name": "Mark Myshatyn",
        "organization": "Los Alamos National Laboratory"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Los Alamos National Laboratory enterprise AI architect Mark Myshatyn explains how agentic AI can accelerate national-security science, including inertial confinement fusion design and simulations on high-performance computing infrastructure. He describes the Venado supercomputer, partnerships with OpenAI, NVIDIA, HPE, and academia, and the challenges of deploying AI under federal governance, procurement, security, isolation, and FedRAMP requirements. He closes by connecting the laboratory’s broader scientific mission to the ChemCam instrument on Mars."
  },
  {
    "slug": "gpt-web-app-generator-10-000-apps-created-in-a-month-matija-sosic",
    "title": "GPT Web App Generator - 10,000 apps created in a month: Matija Sosic",
    "url": "https://ai.engineer/talks/gpt-web-app-generator-10-000-apps-created-in-a-month-matija-sosic",
    "videoId": "atwSCxWFVMg",
    "event": "AI Engineer Summit 2023",
    "durationMs": 563000,
    "speakers": [
      {
        "slug": "matija-sosic",
        "name": "Matija Sosic",
        "organization": "Wasp"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Wasp co-founder Matija Sosic demonstrates Mage, a GPT-powered generator that turns a short description into a downloadable full-stack React, Node.js, Prisma, and Wasp application. He shows a working task-management app, database inspection, and Fly.io deployment, then explains how declarative scaffolding, heuristic error correction, GPT-4 planning, and cheaper GPT-3.5 implementation improve reliability and reduce generation costs."
  },
  {
    "slug": "gpu-cloud-deployment-without-leaving-your-ide-audry-hsu-runpod",
    "title": "GPU Cloud Deployment Without Leaving Your IDE — Audry Hsu, RunPod",
    "url": "https://ai.engineer/talks/gpu-cloud-deployment-without-leaving-your-ide-audry-hsu-runpod",
    "videoId": "zDGHt0LB-dA",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1219000,
    "speakers": [
      {
        "slug": "audry-hsu",
        "name": "Audry Hsu",
        "organization": "RunPod"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "RunPod's Audry Hsu introduces GPU infrastructure options including Pods, Serverless, clusters, and preconfigured model deployments, then demonstrates Flash, a Python SDK that sends decorated functions to cloud GPU workers while keeping surrounding development code local. She contrasts this approach with repeated GitHub, Docker, registry, and GPU-provisioning cycles, and concludes with a multi-model image-composition demonstration using Google's Nano Banana 2."
  },
  {
    "slug": "gpu-less-trust-less-limit-less-reimagining-the-confidential-ai-cloud",
    "title": "GPU-less, Trust-less, Limit-less: Reimagining the Confidential AI Cloud",
    "url": "https://ai.engineer/talks/gpu-less-trust-less-limit-less-reimagining-the-confidential-ai-cloud",
    "videoId": "A0PxE39xaMc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2621000,
    "speakers": [
      {
        "slug": "mike-bursell",
        "name": "Mike Bursell",
        "organization": "Super Protocol"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      }
    ],
    "summary": "Mike Bursell introduces Super Protocol's confidential AI cloud, explaining how hardware-backed trusted execution environments and cryptographic attestation protect sensitive data and proprietary models during training, fine-tuning, and inference. The presentation discusses healthcare data access, model provenance, decentralized AI marketplaces, Intel, AMD, and NVIDIA confidential-computing infrastructure, and demonstrations including an n8n deployment and encrypted confidential-virtual-machine workflows. A second unidentified speaking participant appears in the demonstrations."
  },
  {
    "slug": "graph-intelligence-enhance-reasoning-and-retrieval-using-graph-analytics",
    "title": "Graph Intelligence: Enhance Reasoning and Retrieval Using Graph Analytics",
    "url": "https://ai.engineer/talks/graph-intelligence-enhance-reasoning-and-retrieval-using-graph-analytics",
    "videoId": "GGxAQVbwBL4",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 6076000,
    "speakers": [
      {
        "slug": "andreas-kollegger",
        "name": "Andreas Kollegger",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Neo4j presenters [REDACTED:email_address] Cossette and Andreas Kollegger are listed for a hands-on advanced GraphRAG workshop exploring how graph analytics improve retrieval and reasoning after a knowledge graph has been built. The recording covers LLM-assisted knowledge-graph extraction, workshop and environment setup, Agent Neo example data, document similarity graphs with source provenance, embedding visualization, graph communities, and identifying redundant documents. Audience questions and practical troubleshooting are interspersed throughout."
  },
  {
    "slug": "graphrag-methods-to-create-optimized-llm-context-windows-for-retrieval-jonathan-larson-microsoft",
    "title": "GraphRAG methods to create optimized LLM context windows for Retrieval — Jonathan Larson, Microsoft",
    "url": "https://ai.engineer/talks/graphrag-methods-to-create-optimized-llm-context-windows-for-retrieval-jonathan-larson-microsoft",
    "videoId": "c5qJHr3DnT4",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 909000,
    "speakers": [
      {
        "slug": "jonathan-larson",
        "name": "Jonathan Larson",
        "organization": "Microsoft Research"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Microsoft Research's Jonathan Larson explains how GraphRAG uses structured memory to improve LLM retrieval and repository-level code understanding. Demonstrations contrast ordinary RAG with GraphRAG on a small multifile game, translate that game into working Rust, and extend analysis to Doom's substantially larger codebase. He also introduces BenchmarkQED and discusses LazyGraphRAG results against vector-based RAG across different context-window sizes."
  },
  {
    "slug": "grounded-reasoning-systems-for-cloud-architecture",
    "title": "Grounded Reasoning Systems for Cloud Architecture",
    "url": "https://ai.engineer/talks/grounded-reasoning-systems-for-cloud-architecture",
    "videoId": "9mzfioh1Zag",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1566000,
    "speakers": [
      {
        "slug": "iman-makaremi",
        "name": "Iman Makaremi",
        "organization": "Catio"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Catio AI co-founding leader Iman Makaremi explains an architecture copilot that grounds reasoning in architectural graph structure, requirements, and retrieval context. He contrasts semantic vector search with richer contextual reasoning, describes a hierarchical multi-agent system with chief architect, specialized staff architect, and retriever agents, and discusses structured communication, orchestration control flow, recommendation conflict resolution, evaluation, monitoring, and hallucination detection."
  },
  {
    "slug": "gtm-is-you-victoria-melnikova-evil-martians",
    "title": "GTM Is You - Victoria Melnikova, Evil Martians",
    "url": "https://ai.engineer/talks/gtm-is-you-victoria-melnikova-evil-martians",
    "videoId": "G6IlDzj8OjA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 755000,
    "speakers": [
      {
        "slug": "victoria-melnikova",
        "name": "Victoria Melnikova",
        "organization": "Evil Martians"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Victoria Melnikova of Evil Martians argues that developer-tool and AI startup founders are central to their companies’ go-to-market strategy. Drawing on founder interviews and examples involving PlanetScale and Typesense, she discusses personal branding, product-market fit, AI-saturated distribution channels, San Francisco founder networks, outdoor advertising, events, and developing a distinctive founder voice."
  },
  {
    "slug": "guardrails-first-engineering-member-facing-health-ai",
    "title": "Guardrails First: Engineering Member-Facing Health AI",
    "url": "https://ai.engineer/talks/guardrails-first-engineering-member-facing-health-ai",
    "videoId": "YXEqC05WEI0",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1309000,
    "speakers": [
      {
        "slug": "rashi-agrawal",
        "name": "Rashi Agrawal",
        "organization": "Hinge Health"
      }
    ],
    "topics": [
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Hinge Health AI leader Rashi Agrawal explains how to engineer member-facing healthcare AI around architectural safety constraints: remove protected health information at ingestion, isolate production systems, enforce emergency escalation and clinical routing in deterministic code before model execution, and continuously evaluate clinical responses while checking whether automated judges themselves are reliable. Examples of unsafe chatbot guidance and ECRI’s 2026 hazard ranking motivate a guardrails-first approach to regulated healthcare deployment."
  },
  {
    "slug": "guide-verify-solve-the-engineering-discipline-agentic-development-demands",
    "title": "Guide, Verify, Solve: The Engineering Discipline Agentic Development Demands",
    "url": "https://ai.engineer/talks/guide-verify-solve-the-engineering-discipline-agentic-development-demands",
    "videoId": "03l29gJXpCE",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1351000,
    "speakers": [
      {
        "slug": "anirban-chatterjee",
        "name": "Anirban Chatterjee",
        "organization": "Sonar"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Sonar product-marketing leader Anirban Chatterjee argues that agentic software development requires systematic verification rather than relying on short-lived productivity gains or human review alone. He discusses reported Carnegie Mellon findings on increased static-analysis warnings and code complexity, SonarQube-based model evaluation, research on human acceptance of incorrect AI advice, and a guide-verify-solve approach combining agent guardrails, automated verification, and iterative remediation."
  },
  {
    "slug": "hacking-subagents-into-codex-cli-brian-john-betterup",
    "title": "Hacking Subagents Into Codex CLI — Brian John, BetterUp",
    "url": "https://ai.engineer/talks/hacking-subagents-into-codex-cli-brian-john-betterup",
    "videoId": "5eJqXtevlXg",
    "event": "AI Engineer Code 2025",
    "durationMs": 819000,
    "speakers": [
      {
        "slug": "brian-john",
        "name": "Brian John",
        "organization": "BetterUp"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "BetterUp principal full-stack engineer Brian John demonstrates a proof-of-concept approach for adding subagents to Codex CLI: a parent Codex session invokes a wrapper that launches Codex Exec as a child, exchanges prompts and results through files, and returns output over standard output. He examines sandboxed credential access, permissions, external API communication, a Python AgentExecutor, CODEX_HOME configuration, workspace-write/full-auto execution, and practical timeout considerations."
  },
  {
    "slug": "hacking-the-inference-pareto-frontier",
    "title": "Hacking the Inference Pareto Frontier",
    "url": "https://ai.engineer/talks/hacking-the-inference-pareto-frontier",
    "videoId": "Y2qc0UhDSnc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1225000,
    "speakers": [
      {
        "slug": "kyle-kranen",
        "name": "Kyle Kranen",
        "organization": "NVIDIA"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "NVIDIA's Kyle Kranen explains how inference-system design can reshape the Pareto frontier between application quality, latency, throughput, and cost. He introduces NVIDIA Dynamo and describes KV caching, separated prefill and decode workers, request routing, GPU allocation, and agent-workload structure. A Llama 70B example illustrates a claimed throughput improvement of up to two times at fixed latency using sixteen H100 GPUs; the talk emphasizes that results depend on workload characteristics and worker configuration."
  },
  {
    "slug": "hard-won-lessons-from-building-effective-ai-coding-agents",
    "title": "Hard-Won Lessons from Building Effective AI Coding Agents",
    "url": "https://ai.engineer/talks/hard-won-lessons-from-building-effective-ai-coding-agents",
    "videoId": "I8fs4omN1no",
    "event": "AI Engineer Code 2025",
    "durationMs": 858000,
    "speakers": [
      {
        "slug": "nik-pash",
        "name": "Nik Pash",
        "organization": "Cline"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Cline head of AI Nik Pash argues that stronger frontier models increasingly outperform elaborate coding-agent scaffolding, illustrating the point with Terminal-Bench and Terminus. He contends that meaningful progress comes from difficult benchmarks and reinforcement-learning environments rather than incremental agent tuning, then describes Cline's provisional RL Environments Factory: parallel task qualification, reconstruction of real bugs and fixes, outcome-focused verification, and containerized benchmark generation."
  },
  {
    "slug": "harness-engineering-is-not-enough-why-software-factories-fail",
    "title": "Harness Engineering is not Enough: Why Software Factories Fail",
    "url": "https://ai.engineer/talks/harness-engineering-is-not-enough-why-software-factories-fail",
    "videoId": "Ib5GBkD555M",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1158000,
    "speakers": [
      {
        "slug": "dex-horthy",
        "name": "Dex Horthy",
        "organization": "HumanLayer"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "HumanLayer's Dex Horthy argues that fully autonomous software factories cannot preserve complex codebases through harness engineering alone: coding agents accelerate implementation while inadequate review, incidents, and architectural decay accumulate. He contrasts traditional and agentic development workflows, explains why model-training incentives and weak maintainability benchmarks leave long-term code quality underserved, and advocates human oversight, deliberate architecture, and better verification."
  },
  {
    "slug": "harnessing-the-power-of-llms-locally",
    "title": "Harnessing the Power of LLMs Locally",
    "url": "https://ai.engineer/talks/harnessing-the-power-of-llms-locally",
    "videoId": "MwqUYRQloGw",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1029000,
    "speakers": [
      {
        "slug": "mithun-hunsur",
        "name": "Mithun Hunsur",
        "organization": "Ambient"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Mithun Hunsur of Ambient introduces llm, a Rust library for running large language models locally, and contrasts local inference with cloud-hosted models in terms of model size, hardware, and practical tradeoffs. He explains the project's origins alongside llama.cpp, its unified architecture interface, hardware and platform goals, inference-session memory costs, a Llama 7B demonstration on a MacBook CPU, and applications including data extraction and deployment."
  },
  {
    "slug": "hasura-launch-realtime-data-connectivity-for-ai",
    "title": "Hasura Launch: Realtime Data Connectivity for AI",
    "url": "https://ai.engineer/talks/hasura-launch-realtime-data-connectivity-for-ai",
    "videoId": "himhGiWJXjo",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 434000,
    "speakers": [
      {
        "slug": "tanmai-gopal",
        "name": "Tanmai Gopal",
        "organization": "Hasura"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Hasura cofounder Tanmai Gopal introduces Pacha DDN as a way to connect LLMs securely to live data and business logic across multiple sources. Through a customer-email example, he outlines a unified SQL-like interface for structured data, unstructured data and APIs, schema-aware rules, and autonomous retrieval planning implemented with generated Python code."
  },
  {
    "slug": "healthcare-s-agent-bytecode-x12-as-the-harness-for-ai-agents",
    "title": "Healthcare’s Agent Bytecode: X12 as the Harness for AI Agents",
    "url": "https://ai.engineer/talks/healthcare-s-agent-bytecode-x12-as-the-harness-for-ai-agents",
    "videoId": "UyyOoJmuATU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1225000,
    "speakers": [
      {
        "slug": "vasant-kearney",
        "name": "Vasant Kearney",
        "organization": "Onlay"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Vasant Kearney of Onlay presents X12 healthcare transaction standards as a structured harness for AI agents operating across insurance claims, payer portals, phone calls, imaging, electronic health records, and payment workflows. He describes preserving multimodal context, storing enterprise agent memory in logically separated databases, limiting persistent-memory bias, and revalidating systems when models change. The talk argues that standardized transaction structures can constrain expensive, error-prone multistep reasoning while supporting reliable healthcare revenue-cycle automation."
  },
  {
    "slug": "hiring-building-an-ai-engineering-team",
    "title": "Hiring & Building an AI Engineering Team",
    "url": "https://ai.engineer/talks/hiring-building-an-ai-engineering-team",
    "videoId": "IxXMKT2FDRk",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1747000,
    "speakers": [
      {
        "slug": "dr-bryan-bischof",
        "name": "Dr Bryan Bischof",
        "organization": "Hex"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Hex's Bryan Bischof explains how to define, hire, and build an AI engineering team around practical product delivery. He distinguishes senior software and machine-learning engineering skills from research roles, emphasizes cross-functional collaboration and appropriately sized teams, and discusses infrastructure maturity, domain expertise, security, reskilling, and curiosity during audience questions."
  },
  {
    "slug": "how-agent-o11y-differs-from-traditional-o11y",
    "title": "How agent o11y differs from traditional o11y",
    "url": "https://ai.engineer/talks/how-agent-o11y-differs-from-traditional-o11y",
    "videoId": "XBaznoTRDFI",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1243000,
    "speakers": [
      {
        "slug": "phil-hetzel",
        "name": "Phil Hetzel",
        "organization": "Braintrust"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Braintrust solutions engineering lead Phil Hetzel explains why non-deterministic AI agents require observability beyond uptime, latency, and token metrics. He describes real-time trace analysis, purpose-built storage with SQL access, Rust-based Tantivy search inspired by Apache Lucene, embedding-driven trace clustering, and expert feedback that becomes scalable automated scoring. Audience questions address agent-framework integration and Braintrust’s previous use of ClickHouse."
  },
  {
    "slug": "how-agents-broke-app-level-infrastructure",
    "title": "How agents broke app-level infrastructure",
    "url": "https://ai.engineer/talks/how-agents-broke-app-level-infrastructure",
    "videoId": "_-oIuRH4oGA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 812000,
    "speakers": [
      {
        "slug": "evan-boyle",
        "name": "Evan Boyle",
        "organization": "GenSX"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "GenSX founder Evan Boyle explains why slow, unreliable LLM calls and agent workflows lasting minutes or hours break assumptions inherited from conventional web applications and serverless infrastructure. He compares queues, Airflow, and Temporal; demonstrates background scraping and content-generation experiences; and outlines a React-inspired TypeScript architecture with instrumented OpenAI SDK calls, separate API and compute layers, and Redis Streams."
  },
  {
    "slug": "how-agents-will-unlock-the-500b-promise-of-ai",
    "title": "How agents will unlock the $500B promise of AI",
    "url": "https://ai.engineer/talks/how-agents-will-unlock-the-500b-promise-of-ai",
    "videoId": "Lqq_LcBaJCc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 982000,
    "speakers": [
      {
        "slug": "donald-hruska",
        "name": "Donald Hruska",
        "organization": "Retool"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Retool engineering lead Donald Hruska argues that enterprise AI investment will deliver practical returns when developers build agents with production integrations, security, and guardrails. He connects the success of coding agents and SWE-bench Verified to broader business automation, explains ReAct-style reasoning, tool use, self-verification, and bounded execution loops, and discusses the engineering trade-offs between building proprietary capabilities and buying supporting infrastructure. Audience questions address Retool's own build-versus-buy approach."
  },
  {
    "slug": "how-autoresearch-is-changing-ml-research-zhengyao-jiang-weco-ai",
    "title": "How Autoresearch Is Changing ML Research — Zhengyao Jiang, Weco AI",
    "url": "https://ai.engineer/talks/how-autoresearch-is-changing-ml-research-zhengyao-jiang-weco-ai",
    "videoId": "iCj_ATyThvc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 976000,
    "speakers": [
      {
        "slug": "zhengyao-jiang",
        "name": "Zhengyao Jiang",
        "organization": "Weco AI"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Weco AI cofounder Zhengyao Jiang describes how its autonomous research agent Aiden participated in OpenAI's Parameter Golf challenge, reporting seven leaderboard records across 22 days and approximately 1,300 experiments. He explains Aiden's public-research ingestion, experimentation, pull-request quality gates, and community impact, while arguing that human creativity and agent execution contribute complementary strengths to machine-learning research."
  },
  {
    "slug": "how-blackrock-builds-custom-knowledge-apps-at-scale",
    "title": "How BlackRock Builds Custom Knowledge Apps at Scale",
    "url": "https://ai.engineer/talks/how-blackrock-builds-custom-knowledge-apps-at-scale",
    "videoId": "08mH36_NVos",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1127000,
    "speakers": [
      {
        "slug": "vaibhav-page",
        "name": "Vaibhav Page",
        "organization": "BlackRock"
      },
      {
        "slug": "infant-vasanth",
        "name": "Infant Vasanth",
        "organization": "BlackRock"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "BlackRock engineers Infant Vasanth and Vaibhav Page explain how investment-operations teams can rapidly build custom AI knowledge applications for tasks such as security setup and financial-document extraction. They discuss prompt engineering, evaluation, retrieval-augmented generation versus chain-of-thought strategies, model context limits, deployment scale, and configurable extraction templates, followed by audience questions about data pipelines and information-extraction safeguards."
  },
  {
    "slug": "how-building-with-ai-can-double-the-throughput-of-your-engineering-team",
    "title": "How Building with AI Can Double the Throughput of Your Engineering Team",
    "url": "https://ai.engineer/talks/how-building-with-ai-can-double-the-throughput-of-your-engineering-team",
    "videoId": "4_VQBbs2iQA",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1309000,
    "speakers": [
      {
        "slug": "brian-scanlan",
        "name": "Brian Scanlan",
        "organization": "Intercom"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Intercom senior principal engineer Brian Scanlan explains the company’s effort to double engineering throughput by making AI adoption an organizational expectation and measuring code changes per R&D employee. He describes Intercom’s Fin customer-support agent, its Ruby on Rails platform, experimentation with coding assistants, reusable agent skills and hooks, an agent-first engineering philosophy, and Claude Code-related pull-request automation."
  },
  {
    "slug": "how-claude-code-works",
    "title": "How Claude Code Works",
    "url": "https://ai.engineer/talks/how-claude-code-works",
    "videoId": "RFKCzGlAU6Q",
    "event": "AI Engineer Code 2025",
    "durationMs": 3943000,
    "speakers": [
      {
        "slug": "jared-zoneraich",
        "name": "Jared Zoneraich",
        "organization": "PromptLayer"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "PromptLayer cofounder Jared Zoneraich presents an independent, explicitly non-Anthropic-endorsed workshop examining Claude Code and other coding agents. He emphasizes simple prompt-driven architectures, contrasts prompting with DAG-based orchestration, explains isolated-context sub-agents and tool calls, compares Claude Code with Codex, and demonstrates historical backtesting, LLM assertions, LLM-as-judge evaluation, and headless SDK possibilities."
  },
  {
    "slug": "how-coding-agents-change-software-development-forever-hailong-zhang",
    "title": "How Coding Agents Change Software Development Forever - Hailong Zhang",
    "url": "https://ai.engineer/talks/how-coding-agents-change-software-development-forever-hailong-zhang",
    "videoId": "EUHx5ooJHuQ",
    "event": "AI Engineer Summit 2025",
    "durationMs": 530000,
    "speakers": [
      {
        "slug": "hailong-zhang",
        "name": "Hailong Zhang",
        "organization": "Gru.ai"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Hailong Zhang contrasts synchronous coding assistants such as GitHub Copilot and Cursor with asynchronous agents that autonomously complete repository tasks. He demonstrates Gru.ai’s Test Gru, which detects pull-request changes, generates and runs unit tests, and submits reviewed changes, reporting that more than half of its pull requests were accepted and approximately 80% of tests in its own repository were agent-generated. He then describes evaluation harnesses, selecting among frontier models, fine-tuning GPT-4o, gathering repository context, and sharing agent infrastructure through an Agent OS."
  },
  {
    "slug": "how-deep-research-works",
    "title": "How Deep Research Works",
    "url": "https://ai.engineer/talks/how-deep-research-works",
    "videoId": "eJOjdjO45Sc",
    "event": "AI Engineer Summit 2025",
    "durationMs": 915000,
    "speakers": [
      {
        "slug": "mukund-sridhar",
        "name": "Mukund Sridhar",
        "organization": "Google DeepMind"
      },
      {
        "slug": "aarush-selvan",
        "name": "Aarush Selvan",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Google DeepMind presenters Aarush Selvan and Mukund Sridhar explain how Gemini Deep Research combines increased inference-time compute, web research, user-visible research plans, and iterative agent planning to produce comprehensive reports. They discuss building asynchronous experiences inside a synchronous chatbot, handling failures across long-running multi-service workflows, supporting cross-platform notifications, and managing research context through recency-aware notes stored in RAG. They conclude with a vision for assistants that move beyond information aggregation toward strategic, profession-specific analysis."
  },
  {
    "slug": "how-evals-and-prompts-shape-agent-behavior-preetika-bhateja-daniel-bump-youtube-ads",
    "title": "How Evals and Prompts Shape Agent Behavior — Preetika Bhateja & Daniel Bump, YouTube Ads",
    "url": "https://ai.engineer/talks/how-evals-and-prompts-shape-agent-behavior-preetika-bhateja-daniel-bump-youtube-ads",
    "videoId": "xyL2Ltkh-SA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1169000,
    "speakers": [
      {
        "slug": "preetika-bhateja",
        "name": "Preetika Bhateja",
        "organization": "Google / YouTube Ads"
      },
      {
        "slug": "daniel-bump",
        "name": "Daniel Bump",
        "organization": "Google / YouTube Ads"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Google YouTube Ads presenters Preetika Bhateja and Daniel Bump explain how agent behavior improves through strong LLM-friendly tools, iterative evaluations, and feedback. They describe moving from intuition-led failure analysis to larger golden datasets, calibrating LLM judges against human experts, inspecting agent reasoning, tracking recurring failure patterns, and selecting launch-readiness metrics."
  },
  {
    "slug": "how-fast-are-llm-inference-engines-anyway",
    "title": "How fast are LLM inference engines anyway?",
    "url": "https://ai.engineer/talks/how-fast-are-llm-inference-engines-anyway",
    "videoId": "DeFF3J8T5Pk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 967000,
    "speakers": [
      {
        "slug": "charles-frye",
        "name": "Charles Frye",
        "organization": "Modal"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Modal's Charles Frye examines how increasingly capable open-weight models and modern inference engines make self-hosted LLM deployment practical. He compares the evolving vLLM, SGLang, and TensorRT-LLM landscape, demonstrates Modal's LLM Engineer's Almanac benchmarking interface, and discusses throughput, first-token latency objectives, workload shapes, engine tuning, and numerical-precision tradeoffs."
  },
  {
    "slug": "how-forward-deployed-engineering-is-done-at-cognition",
    "title": "How Forward Deployed Engineering is done at Cognition",
    "url": "https://ai.engineer/talks/how-forward-deployed-engineering-is-done-at-cognition",
    "videoId": "RVxym6mmIns",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1058000,
    "speakers": [
      {
        "slug": "jia-wu",
        "name": "Jia Wu",
        "organization": "Cognition AI"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Cognition deployed engineering lead Jia Wu describes how forward deployed engineers embed with enterprise customers, align Devin and Devin Cloud with high-value business problems, and connect deployment feedback to product development. Wu contrasts token-usage targets with measurable customer outcomes, citing a reported 82% timeline reduction, increased pull-request throughput, and an ETL migration, while emphasizing the role’s combination of technical, business, and customer-facing skills."
  },
  {
    "slug": "how-forward-deployed-engineering-is-done-at-decagon",
    "title": "How Forward Deployed Engineering is done at Decagon",
    "url": "https://ai.engineer/talks/how-forward-deployed-engineering-is-done-at-decagon",
    "videoId": "7wu2hsRfvV0",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1089000,
    "speakers": [
      {
        "slug": "sunny-rekhi",
        "name": "Sunny Rekhi",
        "organization": "Decagon"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Decagon's Sunny Rekhi explains how forward-deployed engineers configure enterprise AI customer-service agents while translating customer requirements into reusable product capabilities. Using support automation, proactive outreach, and CRM integrations as examples, he emphasizes engineering restraint, clear requirements, rapid time-to-value, support-data-driven ROI, and upstreaming bespoke work so every subsequent customer benefits."
  },
  {
    "slug": "how-forward-deployed-engineering-is-done-at-factory",
    "title": "How Forward Deployed Engineering is done at Factory",
    "url": "https://ai.engineer/talks/how-forward-deployed-engineering-is-done-at-factory",
    "videoId": "wpOA-UXynoM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1281000,
    "speakers": [
      {
        "slug": "eno-reyes",
        "name": "Eno Reyes",
        "organization": "Factory"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Factory co-founder and CTO Eno Reyes explains why deployed engineers should connect enterprise customer realities to reusable product development rather than operate as professional-services consultants. He describes a software-factory feedback loop encompassing customer signals, implementation, review, security validation, deployment, and measurable business outcomes. The talk emphasizes agent-ready codebases, deterministic verification, scalable validation, developer experience, and dense reward signals as prerequisites for increasingly autonomous AI engineering and large-scale codebase migrations."
  },
  {
    "slug": "how-forward-deployed-engineering-is-done-at-ramp",
    "title": "How Forward Deployed Engineering is done at Ramp",
    "url": "https://ai.engineer/talks/how-forward-deployed-engineering-is-done-at-ramp",
    "videoId": "ITMXwI6QL6A",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 845000,
    "speakers": [
      {
        "slug": "leo-mehr",
        "name": "Leo Mehr",
        "organization": "Ramp"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Ramp engineering director Leo Mehr outlines two principles for forward deployed engineering: rigorously scope enterprise requests before building, and scale the engineering pipeline with AI agents. Using an SAP S/4HANA integration request and an unnecessary Android implementation as examples, he emphasizes customer context, alternatives, and cross-customer impact. He then describes a Slack-to-Notion request workflow and argues that agents can support context gathering, scoping, specifications, and implementation while human judgment remains essential."
  },
  {
    "slug": "how-forward-deployed-engineering-is-done-redacted-username-kepler",
    "title": "How Forward Deployed Engineering is done Kepler",
    "url": "https://ai.engineer/talks/how-forward-deployed-engineering-is-done-redacted-username-kepler",
    "videoId": "1OMHGsUZiqA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1340000,
    "speakers": [
      {
        "slug": "redacted-username-ganesh",
        "name": "[REDACTED:username] Ganesh",
        "organization": "Kepler"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Kepler co-founder [REDACTED:username] Ganesh argues that forward deployed engineering is fundamentally a product-discovery and product-strategy function, not merely a go-to-market role. Drawing on [REDACTED:username] Foundry, Project Frontline, Citadel, and Kepler, he describes finding customers’ actual problems, replacing over-scoped dashboard projects with rapidly delivered practical solutions, observing workflows on-site, confronting data-infrastructure constraints, and building quick interventions with production durability in mind."
  },
  {
    "slug": "how-google-deepmind-is-researching-the-next-frontier-of-ai-for-gemini-raia-hadsell-vp-of-researc",
    "title": "How Google DeepMind is researching the next Frontier of AI for Gemini — Raia Hadsell, VP of Research",
    "url": "https://ai.engineer/talks/how-google-deepmind-is-researching-the-next-frontier-of-ai-for-gemini-raia-hadsell-vp-of-researc",
    "videoId": "zZsTVBXcbow",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1237000,
    "speakers": [
      {
        "slug": "raia-hadsell",
        "name": "Raia Hadsell",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "Google DeepMind VP of Research Raia Hadsell discusses frontier AI beyond conventional language models, including future Gemini architectures, multimodal embeddings and contrastive retrieval, Matryoshka Representation Learning, GraphCast weather forecasting with spherical graph neural networks, and interactive three-dimensional Genie world models. She also outlines her research background in neural networks, robotics, and game-playing systems."
  },
  {
    "slug": "how-google-deepmind-runs-agents-at-scale-kp-sawhney-ian-ballantyne-google-deepmind",
    "title": "How Google DeepMind Runs Agents at Scale — KP Sawhney & Ian Ballantyne, Google DeepMind",
    "url": "https://ai.engineer/talks/how-google-deepmind-runs-agents-at-scale-kp-sawhney-ian-ballantyne-google-deepmind",
    "videoId": "7gujZrJ9L5I",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1513000,
    "speakers": [
      {
        "slug": "kp-sawhney",
        "name": "KP Sawhney",
        "organization": "Google DeepMind"
      },
      {
        "slug": "ian-ballantyne",
        "name": "Ian Ballantyne",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Google DeepMind engineers Ian Ballantyne and KP Sawhney present an audience-interactive panel on operating AI agents at scale. They demonstrate Google Antigravity's multi-agent development environment, browser and DOM inspection, and reviewable execution traces; discuss Gemini Deep Research through the Interactions API; and field questions about infrastructure monitoring, observability, agent skills versus MCP and CLIs, workflow evaluation, and pull-request review."
  },
  {
    "slug": "how-i-automate-my-own-job-at-hugging-face-using-agents",
    "title": "How I automate my own job at Hugging Face using agents",
    "url": "https://ai.engineer/talks/how-i-automate-my-own-job-at-hugging-face-using-agents",
    "videoId": "FLUoowDJg4I",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1237000,
    "speakers": [
      {
        "slug": "niels-rogge",
        "name": "Niels Rogge",
        "organization": "Hugging Face"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Hugging Face engineer Niels Rogge explains how he automates research-community outreach that encourages authors to migrate models and datasets from scattered hosting services to the Hugging Face Hub. He contrasts a predictable, predefined LLM workflow for identifying papers and opening GitHub issues with a more autonomous follow-up agent built using Anthropic's Claude Agent SDK and deployed with Modal. Reported outcomes include predominantly positive responses, a proposed 400-gigabyte dataset upload, and migration of PaddleOCR models; he also discusses making research benchmarks and technical concepts easier to access."
  },
  {
    "slug": "how-instacart-transformed-its-search-and-discovery-using-an-llm-driven-approach",
    "title": "How Instacart transformed its search and discovery using an LLM-driven approach",
    "url": "https://ai.engineer/talks/how-instacart-transformed-its-search-and-discovery-using-an-llm-driven-approach",
    "videoId": "PjaVHm_3Ljg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1270000,
    "speakers": [
      {
        "slug": "vinesh-gudla",
        "name": "Vinesh Gudla",
        "organization": "Instacart"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Instacart presenters Vinesh Gudla and [REDACTED:username] Tenneti explain how large language models improve grocery search and product discovery beyond engagement-trained, NPMI, and BERT-based approaches. They ground category predictions in taxonomy and conversion data, generate substitute and synonymous query rewrites, address long-tail coverage with a distilled Llama 8B model, and create discovery-oriented search content while managing latency and cost. Audience questions explore natural-language requests and the importance of Instacart-specific context."
  },
  {
    "slug": "how-intuit-uses-llms-to-explain-taxes-to-millions-of-taxpayers",
    "title": "How Intuit uses LLMs to explain taxes to millions of taxpayers",
    "url": "https://ai.engineer/talks/how-intuit-uses-llms-to-explain-taxes-to-millions-of-taxpayers",
    "videoId": "_zl_zimMRak",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1139000,
    "speakers": [
      {
        "slug": "jaspreet-singh",
        "name": "Jaspreet Singh",
        "organization": "Intuit"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Intuit senior staff engineer Jaspreet Singh explains how TurboTax uses its GenOS platform, Intuit Assist, Claude, and OpenAI models to generate personalized tax explanations and answer taxpayer questions at scale. He discusses regulatory and security requirements, RAG and GraphRAG over proprietary tax knowledge, Claude 3 Haiku fine-tuning, tax analysts as prompt engineers, and evaluation using domain experts and LLM-as-a-judge. Audience questions address evaluation accuracy, hybrid retrieval, and complex tax scenarios."
  },
  {
    "slug": "how-juries-and-librarians-can-solve-gtm-s-ai-trust-problem",
    "title": "How Juries and Librarians Can Solve GTM's AI Trust Problem",
    "url": "https://ai.engineer/talks/how-juries-and-librarians-can-solve-gtm-s-ai-trust-problem",
    "videoId": "YZQsWVeN3rE",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1029000,
    "speakers": [
      {
        "slug": "alex-bauer",
        "name": "Alex Bauer",
        "organization": "Upside"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Upside co-founder Alex Bauer explains why agentic go-to-market workflows require clean business data, trustworthy context, and verifiable citations. He describes how Claude helps nontechnical revenue teams build tools, why confident hallucinations undermine business decisions, and how familiar human-team management patterns can make AI agents more reliable."
  },
  {
    "slug": "how-kepler-built-verifiable-ai-for-financial-services",
    "title": "How Kepler Built Verifiable AI for Financial Services",
    "url": "https://ai.engineer/talks/how-kepler-built-verifiable-ai-for-financial-services",
    "videoId": "Tt2kX2sgQio",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1350000,
    "speakers": [
      {
        "slug": "vinoo-ganesh",
        "name": "Vinoo Ganesh",
        "organization": "Kepler"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Kepler CEO and co-founder Vinoo Ganesh explains why financial institutions cannot rely on evaluations alone to make nondeterministic language-model outputs trustworthy. He describes separating AI reasoning from deterministic data retrieval, computation, verification, and source-level provenance so analysts can audit financial work products against underlying records. The talk contrasts this approach with traditional financial-data platforms, discusses alpha decay and institution-specific assumptions, and concludes with audience questions about provenance."
  },
  {
    "slug": "how-llms-work-for-web-devs-gpt-in-600-lines-of-vanilla-js",
    "title": "How LLMs work for Web Devs: GPT in 600 lines of Vanilla JS",
    "url": "https://ai.engineer/talks/how-llms-work-for-web-devs-gpt-in-600-lines-of-vanilla-js",
    "videoId": "ZuiJjkbX0Og",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 6094000,
    "speakers": [
      {
        "slug": "ishan-anand",
        "name": "Ishan Anand"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Ishan Anand demystifies large language models for web developers through a roughly 600-line vanilla JavaScript implementation of GPT-2, adapting his earlier Excel-based Spreadsheets-are-all-you-need project. The workshop examines Transformer architecture, token vocabularies and BPE, semantic and positional embeddings, attention and repeated model blocks, neural-network error and backpropagation, and how GPT-2 relates to later model families, with audience questions throughout."
  },
  {
    "slug": "how-lovable-self-improves-every-hour",
    "title": "How Lovable self-improves every hour",
    "url": "https://ai.engineer/talks/how-lovable-self-improves-every-hour",
    "videoId": "KA5kPbdkK2E",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1145000,
    "speakers": [
      {
        "slug": "benjamin-verbeek",
        "name": "Benjamin Verbeek",
        "organization": "Lovable"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Lovable engineer Benjamin Verbeek explains two continuous-improvement loops for its AI coding platform: identifying sessions where nontechnical users get stuck, extracting and clustering successful resolutions, and measuring project completion; and letting agents report platform friction directly to Slack. He illustrates the second loop with a filename-copying bug involving non-breaking spaces and describes progressing toward automatically evaluated fixes."
  },
  {
    "slug": "how-to-add-secure-code-interpreting-in-your-ai-app",
    "title": "How to add secure code interpreting in your AI app",
    "url": "https://ai.engineer/talks/how-to-add-secure-code-interpreting-in-your-ai-app",
    "videoId": "k0VIgKAUkP4",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 6496000,
    "speakers": [
      {
        "slug": "vasek-mlejnsky",
        "name": "Vasek Mlejnsky",
        "organization": "E2B"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "E2B co-founder Vasek Mlejnsky leads a hands-on workshop building a Claude Artifacts-style AI application that executes model-generated code securely. He combines Claude 3.5 Sonnet, the Vercel AI SDK, function calling, and E2B sandboxes; covers API and repository setup, Python data-analysis libraries and execution outputs; and explains how Docker-based sandbox images become isolated Firecracker virtual machines."
  },
  {
    "slug": "how-to-build-agents-that-run-for-hours-without-losing-the-plot",
    "title": "How to Build Agents That Run for Hours (Without Losing the Plot)",
    "url": "https://ai.engineer/talks/how-to-build-agents-that-run-for-hours-without-losing-the-plot",
    "videoId": "mR-WAvEPRwE",
    "event": "AI Engineer Europe 2026",
    "durationMs": 4540000,
    "speakers": [
      {
        "slug": "ash-prabaker",
        "name": "Ash Prabaker",
        "organization": "Anthropic"
      },
      {
        "slug": "andrew-wilson",
        "name": "Andrew Wilson",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Anthropic Applied AI engineers Ash Prabaker and Andrew Wilson explain how to build agent harnesses that sustain autonomous software work for hours. The workshop traces Claude Code and model improvements, including checkpoints, agent teams, and context compaction, before examining generator-evaluator workflows, subjective-quality rubrics, persistent artifacts, and reducing unnecessary scaffolding as models improve. An extended audience discussion addresses evaluator reliability, Playwright MCP, model comparisons, long-lived products, and minimizing human intervention."
  },
  {
    "slug": "how-to-build-ai-agents-that-actually-work",
    "title": "How to Build AI Agents that Actually Work",
    "url": "https://ai.engineer/talks/how-to-build-ai-agents-that-actually-work",
    "videoId": "7MiFIhlkBoE",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1064000,
    "speakers": [
      {
        "slug": "patrick-dougherty",
        "name": "Patrick Dougherty",
        "organization": "Rasgo"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Rasgo co-founder and former CTO Patrick Dougherty describes practical lessons from rebuilding an enterprise-data product around autonomous AI agents. He defines agents by goal-directed behavior, tool access, and autonomous reasoning; contrasts tool-based retrieval with prompt-injected RAG; and demonstrates how GPT-4o invents an unsupported customer-churn SQL query while o1 recognizes that the Salesforce schema lacks the necessary data. He explains agent-computer interface design, including improved GPT-4o results with JSON and Claude results with XML, discusses Claude 3.5 Sonnet and failure-mode analysis, and cautions against excessive reliance on third-party libraries in production."
  },
  {
    "slug": "how-to-build-an-ai-native-health-company",
    "title": "How to build an AI-Native Health Company",
    "url": "https://ai.engineer/talks/how-to-build-an-ai-native-health-company",
    "videoId": "WJRdLNhrsLQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1039000,
    "speakers": [
      {
        "slug": "dan-feng",
        "name": "Dan Feng",
        "organization": "Maven Clinic"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Dan Feng explains how Maven Clinic is becoming an AI-native healthcare company through internal AI adoption, AI-enabled products, cultural change, and Maven Intelligence, its shared orchestration layer. He describes supporting both Cursor and Claude Code, shifting engineering toward independent AI-assisted implementation, hiring for adaptability, planning around two-to-four-week delivery windows, and improving code-review and release practices to avoid false confidence despite hallucination risks."
  },
  {
    "slug": "how-to-build-an-ai-strategy-that-fails",
    "title": "How To Build an AI Strategy That Fails",
    "url": "https://ai.engineer/talks/how-to-build-an-ai-strategy-that-fails",
    "videoId": "89aQ7T6cMwA",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1023000,
    "speakers": [
      {
        "slug": "hamel-husain",
        "name": "Hamel Husain",
        "organization": "Parlance Labs"
      },
      {
        "slug": "greg-ceccarelli",
        "name": "Greg Ceccarelli",
        "organization": "SpecStory"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Hamel Husain of Parlance Labs and Greg Ceccarelli of SpecStory use an intentionally inverted, satirical presentation to expose enterprise AI strategy failure modes: isolating executives from practitioners, confusing customer value with hype, overspending on GPUs, accumulating technical debt, setting vague goals, pursuing gimmicky applications, drowning teams in documents, and optimizing vanity metrics instead of real outcomes. Their underlying recommendations emphasize cross-functional communication, concrete business problems, realistic costs and timelines, accessible technical practices, and measurements tied to genuine failure modes."
  },
  {
    "slug": "how-to-build-enterprise-aware-agents",
    "title": "How to build Enterprise-aware agents",
    "url": "https://ai.engineer/talks/how-to-build-enterprise-aware-agents",
    "videoId": "hxFpUcvWPcU",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1193000,
    "speakers": [
      {
        "slug": "chau-tran",
        "name": "Chau Tran",
        "organization": "Glean"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Glean engineer Chau Tran contrasts predictable, explicitly orchestrated workflows with flexible LLM-directed agents, examining their respective control, debugging, latency and cost tradeoffs. The talk argues that enterprise agents need organization-specific procedures even as models approach AGI, proposes evaluating execution steps rather than only final outputs, and introduces Workflow Search backed by similar-task retrieval, knowledge graphs and authoritativeness signals."
  },
  {
    "slug": "how-to-build-planning-agents-without-losing-control-yogendra-miraje-factset",
    "title": "How to Build Planning Agents Without Losing Control - Yogendra Miraje, FactSet",
    "url": "https://ai.engineer/talks/how-to-build-planning-agents-without-losing-control-yogendra-miraje-factset",
    "videoId": "sl3icG-IjHo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 958000,
    "speakers": [
      {
        "slug": "yogendra-miraje",
        "name": "Yogendra Miraje",
        "organization": "FactSet"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "FactSet’s Yogendra Miraje explains how to make autonomous planning agents more controllable by distinguishing static workflows from agentic systems, moving beyond reactive agents, and adapting the LLMCompiler plan-and-execute architecture. His implementation represents a blueprint generator, planner, executor, and joiner as LangGraph nodes, uses MCP-compatible tools with explicit contracts, and limits planner complexity through task blueprints. An NVIDIA earnings-call preparation example illustrates structured financial research workflows, followed by audience questions about implementation resources."
  },
  {
    "slug": "how-to-build-the-world-s-fastest-voice-bot",
    "title": "How to build the world's fastest voice bot",
    "url": "https://ai.engineer/talks/how-to-build-the-world-s-fastest-voice-bot",
    "videoId": "dRQHikOrH2A",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1238000,
    "speakers": [
      {
        "slug": "kwindla-hultman-kramer",
        "name": "Kwindla Hultman Kramer",
        "organization": "Daily"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Daily CEO Kwindla Hultman Kramer explains how to build low-latency conversational voice agents by combining WebRTC audio transport, fast transcription, phrase endpointing, colocated LLM inference, and voice synthesis. Healthcare-intake and interactive geography demonstrations illustrate production voice-agent behavior and tool calling, and he introduces Pipecat as an open-source orchestration framework."
  },
  {
    "slug": "how-to-build-trustworthy-ai",
    "title": "How to Build Trustworthy AI",
    "url": "https://ai.engineer/talks/how-to-build-trustworthy-ai",
    "videoId": "n6wHJDqlS1I",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1462000,
    "speakers": [
      {
        "slug": "allie-howe",
        "name": "Allie Howe",
        "organization": "Growth Cyber"
      }
    ],
    "topics": [
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Allie Howe of Growth Cyber explains trustworthy AI as the combination of AI security and AI safety, arguing that product, engineering, and security teams share responsibility for preventing misuse, prompt injection, data leakage, and harmful outputs. She contrasts conventional DevSecOps with MLSecOps for environments such as Databricks and Jupyter notebooks, demonstrates Protect AI's open-source ModelScan, discusses RAG-related injection risks and a clinical-trial matching example, and describes runtime guardrails that enforce security, safety, and business objectives."
  },
  {
    "slug": "how-to-build-world-class-ai-products-sarah-sachs-notion-and-carlos-esteban-braintrust",
    "title": "How to build world-class AI products — Sarah Sachs (Notion) and Carlos Esteban (Braintrust)",
    "url": "https://ai.engineer/talks/how-to-build-world-class-ai-products-sarah-sachs-notion-and-carlos-esteban-braintrust",
    "videoId": "6YdPI9YbjbI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 6226000,
    "speakers": [
      {
        "slug": "sarah-sachs",
        "name": "Sarah Sachs",
        "organization": "Notion"
      },
      {
        "slug": "carlos-esteban",
        "name": "Carlos Esteban",
        "organization": "Braintrust"
      },
      {
        "slug": "doug-guthrie",
        "name": "Doug Guthrie",
        "organization": "Braintrust"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "In this hands-on workshop, Notion AI lead Sarah Sachs explains how observability, curated datasets, LLM-as-a-judge evaluations, and separate retrieval assessments support reliable AI products. Braintrust presenters then demonstrate evaluation scoring, Playgrounds and Experiments, OpenAI-based implementation, production logging, human review, and feedback loops that help technical and nontechnical teams iterate together."
  },
  {
    "slug": "how-to-build-your-own-ai-data-center-in-2025",
    "title": "How to Build Your Own AI Data Center in 2025",
    "url": "https://ai.engineer/talks/how-to-build-your-own-ai-data-center-in-2025",
    "videoId": "3j1dHivahFQ",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1380000,
    "speakers": [
      {
        "slug": "paul-gilbert",
        "name": "Paul Gilbert",
        "organization": "Arista Networks"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Arista Networks technical lead Paul Gilbert explains how enterprises design dedicated AI data-center infrastructure, contrasting training and inference GPU requirements, isolated backend fabrics, front-end storage networks, and east-west traffic. He discusses NVIDIA H100, DGX, and HGX systems, simplified BGP routing, RoCEv2 congestion management using PFC and ECN, Arista EOS, GPU-to-switch coordination, and emerging NIC-centric network designs."
  },
  {
    "slug": "how-to-construct-domain-specific-llm-evaluation-systems",
    "title": "How to construct domain-specific LLM evaluation systems.",
    "url": "https://ai.engineer/talks/how-to-construct-domain-specific-llm-evaluation-systems",
    "videoId": "eLXF0VojuSs",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1125000,
    "speakers": [
      {
        "slug": "hamel-husain",
        "name": "Hamel Husain",
        "organization": "Parlance Labs"
      },
      {
        "slug": "emil-sedgh",
        "name": "Emil Sedgh",
        "organization": "Rechat"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Emil Sedgh and Hamel Husain explain how Rechat moved its real-estate AI assistant, Lucy, from an unreliable GPT-3.5 prototype toward production by building domain-specific evaluations. They recommend examining real failure modes, starting with inexpensive unit tests and assertions, running checks through existing CI and Metabase workflows, and using LangSmith where useful. Synthetic real-estate-agent prompts expand test coverage, while repeated prompt-engineering and evaluation cycles improve application reliability and practical agent performance."
  },
  {
    "slug": "how-to-defend-your-sites-from-ai-bots",
    "title": "How to defend your sites from AI bots",
    "url": "https://ai.engineer/talks/how-to-defend-your-sites-from-ai-bots",
    "videoId": "Gi4V8viBGYQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1212000,
    "speakers": [
      {
        "slug": "david-mytton",
        "name": "David Mytton",
        "organization": "Arcjet"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Arcjet founder David Mytton explains how AI crawlers increase website traffic and infrastructure costs, while distinguishing useful search indexing and user-initiated agents from training crawlers and abusive scraping. He describes inspecting HTTP user agents, checking source IP addresses with reverse DNS, deploying open-source proxy defenses, and using emerging HTTP request-signature standards to verify automated clients."
  },
  {
    "slug": "how-to-evaluate-a-model-for-your-use-case",
    "title": "How to evaluate a model for your use case",
    "url": "https://ai.engineer/talks/how-to-evaluate-a-model-for-your-use-case",
    "videoId": "pj_hKFhnJCw",
    "event": "AI Engineer Summit 2023",
    "durationMs": 452000,
    "speakers": [
      {
        "slug": "emmanuel-turlay",
        "name": "Emmanuel Turlay",
        "organization": "Sematic"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Emmanuel Turlay, then CEO of Sematic, explains why generic metrics and benchmarks cannot establish how a language model will perform on a particular application. After reviewing BLEU, ROUGE, GLUE, HellaSwag, TriviaQA, and ARC, he proposes using an LLM to grade model outputs against application-specific criteria. He contrasts GPT-4 with the less costly Flan-T5 and introduces Airtrain for dataset-based comparison of models including Llama 2 and Falcon."
  },
  {
    "slug": "how-to-hire-ai-engineers-when-everyone-is-cheating-with-ai",
    "title": "How to Hire AI Engineers When Everyone Is Cheating With AI",
    "url": "https://ai.engineer/talks/how-to-hire-ai-engineers-when-everyone-is-cheating-with-ai",
    "videoId": "Zqu0VaJw3vo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 405000,
    "speakers": [
      {
        "slug": "beth-glenfield",
        "name": "Beth Glenfield",
        "organization": "DevDay"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Beth Glenfield argues that AI coding assistants and interview-cheating tools have undermined conventional technical interviews, particularly LeetCode-style assessments. She describes DevDay's alternative: realistic workplace simulations where candidates collaborate with AI agents embodying different teammate roles and are assessed on technical judgment, communication, adaptability, mentoring, and business impact. The approach targets companies that cannot compete with large technology firms through high-volume hiring or expensive hiring mistakes."
  },
  {
    "slug": "how-to-improve-your-agents-academic-lit-review",
    "title": "How to Improve Your Agents: Academic Lit Review",
    "url": "https://ai.engineer/talks/how-to-improve-your-agents-academic-lit-review",
    "videoId": "UOsOfLnAX3Y",
    "event": "AI Engineer Summit 2025",
    "durationMs": 2342000,
    "speakers": [
      {
        "slug": "zhou-yu",
        "name": "Zhou Yu",
        "organization": "Arklex AI; Columbia University"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Columbia University professor and Arklex AI co-founder Zhou Yu reviews research on improving AI agents, beginning with perception, actions, autonomy levels, and human oversight. She examines self-improvement for mathematical reasoning, verifier feedback, on-policy training, and test-time compute scaling, then connects conversational planning and simulation with tree search. The talk closes by emphasizing realistic multi-agent benchmarks that account for task completion, efficiency, and security."
  },
  {
    "slug": "how-to-improve-your-vibe-coding-ian-butler",
    "title": "How to Improve your Vibe Coding — Ian Butler",
    "url": "https://ai.engineer/talks/how-to-improve-your-vibe-coding-ian-butler",
    "videoId": "g03m-WFEu1U",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 450000,
    "speakers": [
      {
        "slug": "ian-butler",
        "name": "Ian Butler",
        "organization": "Bismuth"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Bismuth CEO Ian Butler presents findings from the SM-100 benchmark showing that coding agents can miss real bugs while generating substantial false positives and developer alert fatigue. He recommends reasoning models, OWASP-informed and vulnerability-specific prompts, careful context management, and verification that reported bugs were actually fixed. He closes by outlining Bismuth's automated pull requests, vulnerability scanning, reviews, and integrations."
  },
  {
    "slug": "how-to-kill-the-code-review",
    "title": "How to Kill the Code Review",
    "url": "https://ai.engineer/talks/how-to-kill-the-code-review",
    "videoId": "YgEv7IQzGdM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 986000,
    "speakers": [
      {
        "slug": "ankit-jain",
        "name": "Ankit Jain",
        "organization": "Aviator"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Aviator cofounder Ankit Jain argues that AI-generated code has overwhelmed conventional pull-request review and that replacing human readers with AI reviewers misses review's essential role in team alignment and knowledge sharing. He proposes capturing decisions from coding-agent conversations as acceptance criteria, combining them with reusable guardrails learned from recurring review comments, and verifying executable test plans against live previews. The resulting workflow emphasizes human review of intent, behavioral evidence, and architectural judgment, using deterministic checks whenever possible and language models only when necessary."
  },
  {
    "slug": "how-to-leverage-domain-expertise-chris-lovejoy-notius-labs",
    "title": "How to Leverage Domain Expertise — Chris Lovejoy, Notius Labs",
    "url": "https://ai.engineer/talks/how-to-leverage-domain-expertise-chris-lovejoy-notius-labs",
    "videoId": "kfSDc2eVLo4",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1485000,
    "speakers": [
      {
        "slug": "christopher-lovejoy",
        "name": "Christopher Lovejoy",
        "organization": "Notius Labs"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Christopher Lovejoy argues that successful vertical AI products depend more on systematically incorporating domain expertise than on selecting the most sophisticated models. He introduces the Oracle, Evaluator, and Architect framework for assigning ownership of output quality, measurement, and automated improvement; contrasts Granola's meeting-note workflow with Tandem's clinical-scribe use case; and recommends empowering a principal domain expert with clear accountability and participation in product decisions."
  },
  {
    "slug": "how-to-look-at-your-data-what-to-look-for-how-to-measure",
    "title": "How to look at your data; what to look for, how to measure",
    "url": "https://ai.engineer/talks/how-to-look-at-your-data-what-to-look-for-how-to-measure",
    "videoId": "jryZvCuA0Uc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1163000,
    "speakers": [
      {
        "slug": "jeff-huber",
        "name": "Jeff Huber",
        "organization": "Chroma"
      },
      {
        "slug": "jason-liu",
        "name": "Jason Liu"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Jeff Huber and Jason Liu present a two-part approach to improving AI applications by examining both retrieval inputs and application outputs. Huber explains how generated and real queries, application-specific embedding evaluations, and recall@10 can reveal performance differences that broad MTEB rankings obscure, illustrated with a Weights & Biases chatbot and Weave traces. Liu then demonstrates how extracting and segmenting conversation and usage data makes evaluation gaps actionable, referencing Anthropic’s Clio, and concludes with a brief argument for pricing agent services on successful work delivered."
  },
  {
    "slug": "how-to-quantify-ai-roi-in-software-engineering-stanford-study-120k-devs",
    "title": "How to Quantify AI ROI in Software Engineering (Stanford Study / 120k Devs)",
    "url": "https://ai.engineer/talks/how-to-quantify-ai-roi-in-software-engineering-stanford-study-120k-devs",
    "videoId": "JvosMkuNxF8",
    "event": "AI Engineer Code 2025",
    "durationMs": 1000000,
    "speakers": [
      {
        "slug": "yegor-denisov-blanch",
        "name": "Yegor Denisov-Blanch",
        "organization": "Stanford University"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Stanford researcher Yegor Denisov-Blanch presents a framework for evaluating AI-assisted software engineering using historical Git data, matched teams, and machine-learning models trained to reproduce expert code assessments. He argues that usage quality and codebase cleanliness matter more than raw token consumption, warns that technical debt and rejected AI output can erode productivity and trust, and recommends measuring engineering outcomes alongside guardrail metrics. He closes by inviting organizations, especially Cursor Enterprise users, to participate in further research."
  },
  {
    "slug": "how-to-run-evals-at-scale-thinking-beyond-accuracy-or-similarity",
    "title": "How to Run Evals at Scale: Thinking Beyond Accuracy or Similarity",
    "url": "https://ai.engineer/talks/how-to-run-evals-at-scale-thinking-beyond-accuracy-or-similarity",
    "videoId": "coKKKKh8Vns",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 565000,
    "speakers": [
      {
        "slug": "muktesh-mishra",
        "name": "Muktesh Mishra",
        "organization": "Adobe"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Adobe engineer Muktesh Mishra explains how to evaluate nondeterministic AI applications beyond basic accuracy and similarity. He recommends starting with synthetic data, iteratively refining multiple evaluation datasets, tailoring metrics to RAG versus code-generation systems, and scaling evaluation pipelines through caching, orchestration, and parallelism. He concludes by advocating eval-driven development and continuous measurement, monitoring, analysis, and iteration while balancing fidelity against speed."
  },
  {
    "slug": "how-to-secure-agents-using-oauth",
    "title": "How to Secure Agents using OAuth",
    "url": "https://ai.engineer/talks/how-to-secure-agents-using-oauth",
    "videoId": "blmAkayzE8M",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1139000,
    "speakers": [
      {
        "slug": "jared-hanson",
        "name": "Jared Hanson",
        "organization": "Keycard"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Keycard co-founder and Passport.js creator Jared Hanson explains why AI agents should replace broadly scoped, long-lived API keys with OAuth-based delegated access. He introduces OAuth clients, resource servers, authorization servers, consent, authorization-code flows, and access tokens; distinguishes OpenID Connect authentication and signed ID tokens; and applies these patterns to MCP-connected agents. He also examines client-registration friction, URL-based agent identity, edge-device constraints, transactional authorization, finer-grained scopes, and emerging security questions around voice, video, and background agents."
  },
  {
    "slug": "how-to-talk-to-statues-joe-reeve-elevenlabs",
    "title": "How to talk to statues — Joe Reeve, ElevenLabs",
    "url": "https://ai.engineer/talks/how-to-talk-to-statues-joe-reeve-elevenlabs",
    "videoId": "u-rJwPPU3QA",
    "event": "AI Engineer Europe 2026",
    "durationMs": 2008000,
    "speakers": [
      {
        "slug": "joe-reeve",
        "name": "Joe Reeve",
        "organization": "ElevenLabs"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "ElevenLabs growth engineer Joe Reeve demonstrates an application that photographs statues, uses OpenAI to identify and research them, and creates character-specific synthetic voices for interactive conversations. He discusses ElevenLabs voice design, multimodal agents, culturally appropriate accents, conversational interruption, developer-growth experimentation, rapid prototyping, and the role of music in effective product demos. The recording also includes audience participation and prerecorded character voices."
  },
  {
    "slug": "how-to-train-your-agent-building-reliable-agents-with-rl",
    "title": "How to Train Your Agent: Building Reliable Agents with RL",
    "url": "https://ai.engineer/talks/how-to-train-your-agent-building-reliable-agents-with-rl",
    "videoId": "gEDl9C8s_-4",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1188000,
    "speakers": [
      {
        "slug": "kyle-corbitt",
        "name": "Kyle Corbitt",
        "organization": "OpenPipe"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Kyle Corbitt presents OpenPipe’s ART·E email-assistant case study, showing how reinforcement learning improves a Qwen 2.5 14B agent beyond prompted frontier-model baselines. He reports 96% accuracy versus 90% for o3, compares inference costs of $55 and $8 per thousand searches for o3 and o4-mini, and estimates approximately $80 of GPU time plus one engineering week for training. The talk also examines reward design, verifiable rewards, lower-error agent behavior, and reward-hacking failure modes."
  },
  {
    "slug": "how-transformers-finally-ate-vision",
    "title": "How Transformers Finally Ate Vision",
    "url": "https://ai.engineer/talks/how-transformers-finally-ate-vision",
    "videoId": "VhfAVA3BG2I",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1025000,
    "speakers": [
      {
        "slug": "isaac-robinson",
        "name": "Isaac Robinson",
        "organization": "Roboflow"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Roboflow's Isaac Robinson explains how vision transformers displaced convolutional architectures through scalable pretraining, learned inductive biases, and infrastructure inherited from large language models. Tracing Swin, ConvNeXt, Hiera, masked autoencoders, DINO models, and FlashAttention, he connects backbone design to segmentation foundation models and practical deployment tradeoffs, including Roboflow-reported speed comparisons against fine-tuned SAM3."
  },
  {
    "slug": "how-we-built-zeta2-training-an-edit-prediction-model-in-production-ben-kunkle-zed",
    "title": "How We Built Zeta2: Training an Edit Prediction Model in Production — Ben Kunkle, Zed",
    "url": "https://ai.engineer/talks/how-we-built-zeta2-training-an-edit-prediction-model-in-production-ben-kunkle-zed",
    "videoId": "phchDt63qAA",
    "event": "AI Engineer Europe 2026",
    "durationMs": 650000,
    "speakers": [
      {
        "slug": "ben-kunkle",
        "name": "Ben Kunkle",
        "organization": "Zed"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Zed edit-predictions lead Ben Kunkle explains how Zeta2 predicts code edits from cursor context, recent edits, definitions, and diagnostics while meeting per-keystroke latency requirements. He describes teacher-student distillation, repairing teacher outputs, JSONL-based training infrastructure, settled-state capture, Levenshtein-distance evaluation, and using an increasingly capable student to reduce expensive teacher sampling. The talk closes with production experiment allocation, acceptance-rate and latency monitoring, a Seed-Coder-based release, and an audience question about settled-state signals."
  },
  {
    "slug": "how-we-hacked-yc-spring-2025-batch-s-ai-agents",
    "title": "How we hacked YC Spring 2025 batch’s AI agents",
    "url": "https://ai.engineer/talks/how-we-hacked-yc-spring-2025-batch-s-ai-agents",
    "videoId": "kv-QAuKWllQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1053000,
    "speakers": [
      {
        "slug": "rene-brandel",
        "name": "Rene Brandel",
        "organization": "Casco"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Casco founder and CEO Rene Brandel explains how red-teaming Y Combinator AI-agent startups exposed security weaknesses beyond prompt injection, including insecure direct object references, cross-user data access, network discovery, and access to customer data through BigQuery. He recommends explicit authentication and authorization, ownership-aware access controls, and row-level security, then answers questions about system-prompt extraction and isolation."
  },
  {
    "slug": "how-we-scaled-500m-ai-agents-in-production-with-2-engineers",
    "title": "How we scaled 500m AI agents in production with 2 engineers",
    "url": "https://ai.engineer/talks/how-we-scaled-500m-ai-agents-in-production-with-2-engineers",
    "videoId": "zM9RYqCcioM",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1124000,
    "speakers": [
      {
        "slug": "mustafa-ali",
        "name": "Mustafa Ali",
        "organization": "Method Financial"
      },
      {
        "slug": "kyle-corbitt",
        "name": "Kyle Corbitt",
        "organization": "OpenPipe"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Mustafa Ali of Method Financial and Kyle Corbitt of OpenPipe explain how Method scaled financial AI agents processing unstructured liability data. An initial GPT-4 workflow proved effective but cost $70,000 in its first production month and encountered prompt-engineering, caching, and scaling limitations. They discuss deploying fine-tuned open-source models, comparing model error rates, validating financial outputs, and balancing cost, accuracy, and latency across production workloads."
  },
  {
    "slug": "how-we-solved-context-management-in-agents-sally-ann-delucia",
    "title": "How We Solved Context Management in Agents — Sally-Ann DeLucia",
    "url": "https://ai.engineer/talks/how-we-solved-context-management-in-agents-sally-ann-delucia",
    "videoId": "esY99nYXxR4",
    "event": "AI Engineer Europe 2026",
    "durationMs": 977000,
    "speakers": [
      {
        "slug": "sally-ann-delucia",
        "name": "Sally-Ann DeLucia",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Arize product leader Sally-Ann DeLucia explains how building Alyx, an AI engineering agent operating on observability traces and spans, exposed escalating context-window and long-session problems. She describes why unconstrained LLM summarization was unreliable, frames context management as both an engineering and product-design challenge, and discusses retrieving prior conversation context from database-backed long-term memory. An audience question addresses Claude Code and cache invalidation."
  },
  {
    "slug": "how-we-taught-agents-to-use-good-retrieval-hanna-lichtenberg-mixedbread-ai",
    "title": "How we taught agents to use good retrieval - Hanna Lichtenberg, Mixedbread AI",
    "url": "https://ai.engineer/talks/how-we-taught-agents-to-use-good-retrieval-hanna-lichtenberg-mixedbread-ai",
    "videoId": "1IdzkRVmWAA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 868000,
    "speakers": [
      {
        "slug": "hanna-lichtenberg",
        "name": "Hanna Lichtenberg",
        "organization": "Mixedbread"
      },
      {
        "slug": "aamir-shakir",
        "name": "Aamir Shakir",
        "organization": "Mixedbread"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Hanna Lichtenberg and a second Mixedbread presenter identified as Aamir explain how weak retrieval tools create an oracle gap between capable language models and their real-world research performance. Using BrowseComp+ and OfficeQA Pro, they contrast keyword-oriented search with Mixedbread’s semantic and late-interaction retrieval, then describe an agent harness combining grep, semantic search, parallel search rounds, explicit evidence planning, supervised fine-tuning, and on-policy reinforcement learning. They conclude with preliminary benchmark results and a production Mixedbread Agentic Search offering."
  },
  {
    "slug": "how-web-data-infrastructure-powers-the-next-generation-of-ai",
    "title": "How Web Data Infrastructure Powers the Next Generation of AI",
    "url": "https://ai.engineer/talks/how-web-data-infrastructure-powers-the-next-generation-of-ai",
    "videoId": "1UmZHb_E_SM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1143000,
    "speakers": [
      {
        "slug": "patricija-zemaityte",
        "name": "Patricija Žemaitytė",
        "organization": "Oxylabs"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Oxylabs product manager Patricija Žemaitytė explains why production AI depends on continuously adapting infrastructure for fresh public web data. She traces the evolution of a multimodal video API, describes redesigning SERP retrieval from a four-second baseline to a reported 550-millisecond Fast Search API, and examines proxy-backed Web Unblocker scaling, load testing, and the gap between development success and real-world reliability."
  },
  {
    "slug": "how-zapier-builds-ai-products-and-features-with-the-help-of-braintrust",
    "title": "How Zapier Builds AI Products and Features With the Help of Braintrust",
    "url": "https://ai.engineer/talks/how-zapier-builds-ai-products-and-features-with-the-help-of-braintrust",
    "videoId": "fOtTHWeU6B8",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 899000,
    "speakers": [
      {
        "slug": "ankur-goyal",
        "name": "Ankur Goyal",
        "organization": "Braintrust"
      },
      {
        "slug": "olmo-maldonado",
        "name": "Olmo Maldonado",
        "organization": "Zapier"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Zapier’s Olmo Maldonado and Braintrust’s Ankur Goyal explain how AI Zap Builder and Zapier Copilot were developed around systematic evaluations and observability. They describe synthetic evaluation data, CI and manual test runs, provider load testing, logic-based and LLM-based graders, and rapid regression detection. The presentation also contrasts GPT-4 Turbo’s accuracy-versus-latency tradeoff with a reported reduction in streaming copilot response time from roughly 14 seconds to three seconds after adopting GPT-4o."
  },
  {
    "slug": "html-is-all-agents-need",
    "title": "HTML Is All Agents Need",
    "url": "https://ai.engineer/talks/html-is-all-agents-need",
    "videoId": "Cz4v1WHVyZc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 913000,
    "speakers": [
      {
        "slug": "james-russo",
        "name": "James Russo",
        "organization": "HeyGen"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "HeyGen engineer and HyperFrames co-creator James Russo explains why AI agents can generate more flexible videos by authoring familiar HTML, CSS, and JavaScript instead of learning restrictive video-specific frameworks. He contrasts HyperFrames with traditional creative tools and Remotion, describes designing for Gemini 3 Flash and handling asynchronous browser rendering, and emphasizes creative taste, iterative workflows, human editing, and an open-source framework."
  },
  {
    "slug": "html-is-all-you-need-for-agents-to-make-graphics",
    "title": "HTML is All You Need (for Agents to Make Graphics)",
    "url": "https://ai.engineer/talks/html-is-all-you-need-for-agents-to-make-graphics",
    "videoId": "JRTAtZ5iBkU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 420000,
    "speakers": [
      {
        "slug": "amol-kapoor",
        "name": "Amol Kapoor",
        "organization": "Nori"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Amol Kapoor of Nori argues that AI agents struggle with graphics chiefly when forced into human-oriented canvases or coordinate-heavy SVG workflows. Using ARC-AGI and Simon Willison’s pelican-on-a-bicycle test as context, he proposes HTML as an agent-native, structured representation that browsers can render into polished slides, documents, videos, and PDFs. He describes Nori’s use of HTML and CSS for branded visual assets and positions Nori Sessions as a company-context-aware workflow for producing them."
  },
  {
    "slug": "human-seeded-evals-samuel-colvin-pydantic",
    "title": "Human seeded Evals — Samuel Colvin, Pydantic",
    "url": "https://ai.engineer/talks/human-seeded-evals-samuel-colvin-pydantic",
    "videoId": "o_LRtAomJCs",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 722000,
    "speakers": [
      {
        "slug": "samuel-colvin",
        "name": "Samuel Colvin",
        "organization": "Pydantic"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Samuel Colvin explains why type safety matters when building and refactoring AI applications, then demonstrates Pydantic AI agent loops, structured extraction, schema validation, validation-driven retries, and typed dependencies. Using Gemini Flash and Pydantic Logfire, he shows agent tracing, model-call visibility, timing, and cost reporting. Despite the advertised title, he explicitly states that evaluations will not be covered."
  },
  {
    "slug": "hybridrag-a-fusion-of-graph-and-vector-retrieval-to-enhance-data-interpretation",
    "title": "HybridRAG: A Fusion of Graph and Vector Retrieval to Enhance Data Interpretation",
    "url": "https://ai.engineer/talks/hybridrag-a-fusion-of-graph-and-vector-retrieval-to-enhance-data-interpretation",
    "videoId": "-tgQa8Fzf80",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1224000,
    "speakers": [
      {
        "slug": "mitesh-patel",
        "name": "Mitesh Patel",
        "organization": "NVIDIA"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "finance",
        "name": "Finance"
      }
    ],
    "summary": "NVIDIA developer advocate manager Mitesh Patel presents HybridRAG, combining knowledge-graph relationships with vector-based retrieval to improve question answering over complex documents. He explains entity-relationship triplets and hybrid pipeline construction through an ExxonMobil financial-document example, discusses GPU-accelerated graph operations using cuGraph and NetworkX, and describes evaluating retrieval workflows with Ragas and improving triplet extraction through data cleaning."
  },
  {
    "slug": "hypermode-launch",
    "title": "Hypermode Launch",
    "url": "https://ai.engineer/talks/hypermode-launch",
    "videoId": "8k848OeLc9E",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 303000,
    "speakers": [
      {
        "slug": "kevin-van-gundy",
        "name": "Kevin Van Gundy",
        "organization": "Hypermode"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Kevin Van Gundy introduces Hypermode, drawing on his experience at Vercel to argue that rapid iteration and incremental adoption matter more than committing early to rigid architectures. He presents Hypermode as an AI runtime and surrounding developer tools that integrate models and data, support interchangeable models and inference observability, provide editor typeahead, and streamline retrieval-augmented generation through integrated embedding and search. He closes by inviting attendees to a practical workshop covering natural-language search, intelligent sorting, and outlier detection."
  },
  {
    "slug": "hyperspace-more-nodes-is-all-you-need",
    "title": "Hyperspace: More Nodes Is All You Need",
    "url": "https://ai.engineer/talks/hyperspace-more-nodes-is-all-you-need",
    "videoId": "-zzP0EUsZz4",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 348000,
    "speakers": [
      {
        "slug": "nicolas-schlaepfer",
        "name": "Nicolas Schlaepfer",
        "organization": "Hyperspace"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Nicolas Schlaepfer introduces Hyperspace's decentralized AI inference network and its AI OS desktop product, then demonstrates an agentic application that converts user requests into editable directed acyclic graphs. Its HyperEngine V3 orchestration model streams workflow nodes into a React Flow editor, while Puppeteer and Beautiful Soup prepare web content for language models. The demonstration combines Qwen2 Instruct for reasoning, Llama 3 70B for summarization, and executable Python outputs."
  },
  {
    "slug": "i-gave-an-ai-agent-the-keys-to-my-life-here-s-what-happened",
    "title": "I Gave an AI Agent the Keys to My Life (Here's What Happened)",
    "url": "https://ai.engineer/talks/i-gave-an-ai-agent-the-keys-to-my-life-here-s-what-happened",
    "videoId": "sJ2jc7leKBk",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1174000,
    "speakers": [
      {
        "slug": "radek-sienkiewicz",
        "name": "Radek Sienkiewicz",
        "organization": "OpenClaw"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "OpenClaw maintainer Radek Sienkiewicz describes incrementally giving a personal AI agent access to his communications, files, calendar, and operating system. His setup combines WhatsApp, Telegram, and Discord interfaces with an approximately 3,000-note Obsidian knowledge base, QMD search, contextual bookmark enrichment, overnight index maintenance, automated recovery, and LLM-assisted workflows designed to reduce work for his future self."
  },
  {
    "slug": "i-run-a-fleet-of-ai-agents-across-three-machines-here-s-what-broke",
    "title": "I Run a Fleet of AI Agents Across Three Machines. Here's What Broke.",
    "url": "https://ai.engineer/talks/i-run-a-fleet-of-ai-agents-across-three-machines-here-s-what-broke",
    "videoId": "4kYl2_mqmnQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 551000,
    "speakers": [
      {
        "slug": "kyle-jaejun-lee",
        "name": "Kyle Jaejun Lee",
        "organization": "KRAFTON"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Kyle Jaejun Lee describes operating a personal fleet of AI coding agents across a MacBook and two always-on Linux machines. He replaces manual coordination with CEO, VP, manager, and worker agent roles; persists agent state and handoffs on disk; and resets Claude contexts instead of relying on compaction. Operational failures include overcrowded tmux panes, Claude Code and MCP memory exhaustion, isolated-environment problems, laptop interruptions, and divergent cross-machine state. His evolving architecture transfers work with Git, SSH, and tmux, centralizes reviews on an always-on Linux gateway, and leaves consistency, secure credential handoffs, local-tool access, and resource management unresolved."
  },
  {
    "slug": "i-ve-never-seen-anything-scarier-than-an-llm-with-tool-calls-erik-meijer-aka-headinthebox",
    "title": "\"I've never seen anything scarier than an LLM with tool calls.\" — Erik Meijer aka @HeadinTheBox",
    "url": "https://ai.engineer/talks/i-ve-never-seen-anything-scarier-than-an-llm-with-tool-calls-erik-meijer-aka-headinthebox",
    "videoId": "-CnA2lGfymY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1273000,
    "speakers": [
      {
        "slug": "erik-meijer",
        "name": "Erik Meijer",
        "organization": "Leibniz Labs"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Erik Meijer of Leibniz Labs argues that tool-enabled AI agents are unsafe when prompt injection and unchecked side effects allow models to act without enforceable guarantees. Using Lean, Dafny, elementary type systems, and formal verification, he outlines an air-gapped execution architecture in which agents propose plans accompanied by machine-checkable safety proofs before actions execute."
  },
  {
    "slug": "identity-for-ai-agents",
    "title": "Identity for AI Agents",
    "url": "https://ai.engineer/talks/identity-for-ai-agents",
    "videoId": "VSdV-AdSlis",
    "event": "AI Engineer Code 2025",
    "durationMs": 4932000,
    "speakers": [],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "[REDACTED:username] presenters Patrick Riley and Carlos Galan lead a hands-on workshop on securing AI agents and MCP-connected applications. They explain asynchronous human approvals, Token Vault delegated access and token exchange, fine-grained authorization, and AI security risks before demonstrating a chatbot, trading tools, connected accounts, custom API clients, and an MCP implementation."
  },
  {
    "slug": "imagination-engineering",
    "title": "Imagination Engineering",
    "url": "https://ai.engineer/talks/imagination-engineering",
    "videoId": "Z2Erdirpudo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 964000,
    "speakers": [
      {
        "slug": "eve-bouffard",
        "name": "Eve Bouffard",
        "organization": "Y Combinator"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Y Combinator design leader Eve Bouffard argues that increasingly capable AI makes imagination, rather than implementation, the central constraint on building new things. She demonstrates capturing ideas in a Slack channel, transforming them into an AI-generated personal website, and rapidly exploring historical thinkers through custom software, while discussing Paul Graham's influence, design systems, shaders, and agent-readable documentation."
  },
  {
    "slug": "improving-agents-is-a-data-mining-problem",
    "title": "Improving Agents is a Data Mining Problem",
    "url": "https://ai.engineer/talks/improving-agents-is-a-data-mining-problem",
    "videoId": "CvRngaQZQ3Y",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1202000,
    "speakers": [
      {
        "slug": "vivek-trivedy",
        "name": "Vivek Trivedy",
        "organization": "LangChain"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "LangChain's Vivek Trivedy argues that improving autonomous agents is fundamentally a data-mining problem: capture execution traces, inspect failures and context-compaction effects, compare model behavior, and turn production data into repeatable improvement loops. He discusses scaling trace analysis with cheaper open models, LangSmith Engine, distillation and supervised fine-tuning, reinforcement-learning approaches, and continual learning."
  },
  {
    "slug": "in-the-land-of-ai-agents-the-verifiers-are-king",
    "title": "In the Land of AI Agents, the Verifiers Are King",
    "url": "https://ai.engineer/talks/in-the-land-of-ai-agents-the-verifiers-are-king",
    "videoId": "VrpEyglYgeU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1133000,
    "speakers": [
      {
        "slug": "tariq-shaukat",
        "name": "Tariq Shaukat",
        "organization": "Sonar"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Sonar CEO Tariq Shaukat argues that enterprise value from AI coding agents depends less on generation speed than on rigorous verification. Drawing on METR task-horizon benchmarks and concerns about unreliable outputs, code complexity, and fading productivity gains, he presents Sonar's Agent Centric Development Cycle: Guide agents with architectural context, Verify their outputs through layered checks and quality gates, and Solve accumulated maintenance problems. He introduces Sonar Vortex and advocates integrating verification, CI, and technical-debt remediation into a continuously improving agentic software development lifecycle."
  },
  {
    "slug": "information-retrieval-from-the-ground-up",
    "title": "Information Retrieval from the Ground Up",
    "url": "https://ai.engineer/talks/information-retrieval-from-the-ground-up",
    "videoId": "4Xe_iMYxBQc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 6487000,
    "speakers": [
      {
        "slug": "philipp-krenn",
        "name": "Philipp Krenn",
        "organization": "Elastic"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Elastic's Philipp Krenn presents a hands-on workshop on the retrieval foundations behind RAG, using Elasticsearch and Apache Lucene to explain keyword, vector, and hybrid search. Demonstrations cover language-aware analysis, tokenization, n-grams, document indexing, scoring, learned token representations, and matching-fragment highlighting. Audience discussion considers pgvector and change-data-capture tradeoffs, and the workshop demonstrates semantic reranking with Elastic Rerank."
  },
  {
    "slug": "infra-behind-krea-2-how-to-train-and-serve-at-scale",
    "title": "Infra behind Krea 2 - How to train and serve at scale",
    "url": "https://ai.engineer/talks/infra-behind-krea-2-how-to-train-and-serve-at-scale",
    "videoId": "byn9PURoBNY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1015000,
    "speakers": [
      {
        "slug": "gabriel-jorge-menezes",
        "name": "Gabriel Jorge Menezes",
        "organization": "Krea.ai"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Gabriel Jorge Menezes explains how Krea trained its Krea 2 image model from scratch across large InfiniBand-connected GPU clusters, adapting ideas from language-model research to diffusion transformers. He details operational observability, GPU thermal throttling, misleading utilization figures, tensor-core and networking metrics, then describes Kueue gang scheduling, Kubernetes workload priorities, and practical infrastructure for serving accelerated image-generation models."
  },
  {
    "slug": "infra-that-fixes-itself-thanks-to-coding-agents-mahmoud-abdelwahab-railway",
    "title": "Infra that fixes itself, thanks to coding agents — Mahmoud Abdelwahab, Railway",
    "url": "https://ai.engineer/talks/infra-that-fixes-itself-thanks-to-coding-agents-mahmoud-abdelwahab-railway",
    "videoId": "Q5IVm_CxN2w",
    "event": "AI Engineer Code 2025",
    "durationMs": 1088000,
    "speakers": [
      {
        "slug": "mahmoud-abdelwahab",
        "name": "Mahmoud Abdelwahab",
        "organization": "Railway"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Mahmoud Abdelwahab demonstrates railway-autofix, a self-healing infrastructure prototype that monitors Railway deployments for memory leaks, elevated HTTP errors, latency, and resource anomalies, then uses a coding agent to prepare corrective GitHub pull requests. The architecture combines scheduled Railway API metric collection, Inngest durable workflows that avoid repeating completed work, and a Railway-hosted OpenCode headless agent server implemented with Bun."
  },
  {
    "slug": "insights-from-snorkel-ai-running-azure-ai-infrastructure",
    "title": "Insights from Snorkel AI running Azure AI Infrastructure",
    "url": "https://ai.engineer/talks/insights-from-snorkel-ai-running-azure-ai-infrastructure",
    "videoId": "LJa1SjCkYas",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1246000,
    "speakers": [
      {
        "slug": "lachlan-ainley",
        "name": "Lachlan Ainley",
        "organization": "Microsoft"
      },
      {
        "slug": "humza-iqbal",
        "name": "Humza Iqbal",
        "organization": "Snorkel"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Snorkel researcher Humza Iqbal and Microsoft representative Lachlan Ainley discuss why enterprise AI requires curated, programmatic data development beyond off-the-shelf models and basic RAG. They cover subject-matter-expert feedback, domain-specific evaluation, long-context testing, alignment, synthetic vision-language data, and practical deployment on Azure AI infrastructure, including distributed GPU training and migration from NVIDIA A100 to H100 accelerators."
  },
  {
    "slug": "insights-on-building-ai-teams",
    "title": "Insights on Building AI teams",
    "url": "https://ai.engineer/talks/insights-on-building-ai-teams",
    "videoId": "q_ixa5EW8DY",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1230000,
    "speakers": [
      {
        "slug": "heath-black",
        "name": "Heath Black",
        "organization": "SignalFire"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "SignalFire’s Heath Black explains how its Beacon AI talent-intelligence platform informs AI-team hiring. He argues for filtering candidates by demonstrated engineering experience, open-source work, applied ML and MLOps skills rather than relying primarily on elite-school or PhD credentials; tracking talent movement across leading AI companies and geographic hubs; timing outreach around likely job changes; and understanding candidates’ career narratives."
  },
  {
    "slug": "intelligence-continual-learning-expertise",
    "title": "Intelligence + Continual Learning = Expertise",
    "url": "https://ai.engineer/talks/intelligence-continual-learning-expertise",
    "videoId": "I6aiEf3aEFQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1183000,
    "speakers": [
      {
        "slug": "yu-su",
        "name": "Yu Su",
        "organization": "NeoCognition"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Yu Su, a professor at The Ohio State University and CEO of NeoCognition, argues that general intelligence and situated expertise are different capabilities. Although multimodal language agents excel in structured coding environments, their brittleness, inefficient token use, and difficulty with computer use reflect missing domain-specific knowledge rather than intelligence alone. He defines continual learning as adaptive compression of experience into reusable structures, proposes that effective learning can produce unbounded expertise from bounded intelligence, and describes company-specific human-AI learning loops that accumulate institutional memory."
  },
  {
    "slug": "intro-to-graphrag",
    "title": "Intro to GraphRAG",
    "url": "https://ai.engineer/talks/intro-to-graphrag",
    "videoId": "J-9EbJBxcbg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 4715000,
    "speakers": [
      {
        "slug": "zach-blumenfeld",
        "name": "Zach Blumenfeld",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Neo4j product specialist Zach Blumenfeld leads a hands-on introductory GraphRAG workshop covering hosted Jupyter setup, knowledge graph construction from structured and unstructured data, Cypher queries, graph pattern matching, vector search, embeddings, and community detection. Exercises explore resume-derived graph data and conclude with references to Neo4j MCP servers, ADK examples, and further knowledge graph construction."
  },
  {
    "slug": "introducing-strands-agents-an-open-source-ai-agents-sdk",
    "title": "Introducing Strands Agents, an Open Source AI Agents SDK",
    "url": "https://ai.engineer/talks/introducing-strands-agents-an-open-source-ai-agents-sdk",
    "videoId": "Q3NreEAdKMc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 866000,
    "speakers": [
      {
        "slug": "suman-debnath",
        "name": "Suman Debnath",
        "organization": "Amazon Web Services (AWS)"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Suman Debnath introduces AWS-backed Strands Agents, an open-source, model-driven SDK that builds AI agents around models and tools instead of extensive orchestration. He demonstrates built-in tools and Amazon Bedrock defaults alongside LiteLLM and Ollama alternatives, then connects a Strands agent to an MCP server for generating Manim-based mathematical animations. He closes by inviting contributions to the project’s public GitHub repositories."
  },
  {
    "slug": "introduction-to-llm-serving-with-sglang",
    "title": "Introduction to LLM serving with SGLang",
    "url": "https://ai.engineer/talks/introduction-to-llm-serving-with-sglang",
    "videoId": "Ahtaha9fEM0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2622000,
    "speakers": [
      {
        "slug": "philip-kiely",
        "name": "Philip Kiely",
        "organization": "Baseten"
      },
      {
        "slug": "yineng-zhang",
        "name": "Yineng Zhang",
        "organization": "Baseten"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Baseten presenters Philip Kiely and Yineng Zhang introduce SGLang as an open-source serving framework for language and multimodal models, guide attendees through workshop setup and model deployment, and discuss GPU-based inference optimization, quantization, hardware-specific configuration, CUDA kernels, and cache-aware routing. They also explain how to contribute to the SGLang repository and community before taking audience questions about configuration and security."
  },
  {
    "slug": "invisible-users-invisible-interfaces-accelerating-design-iteration-with-ai-simulation",
    "title": "Invisible Users, Invisible Interfaces: Accelerating Design Iteration with AI Simulation",
    "url": "https://ai.engineer/talks/invisible-users-invisible-interfaces-accelerating-design-iteration-with-ai-simulation",
    "videoId": "8D_VdU6DBhI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 757000,
    "speakers": [
      {
        "slug": "alex-liss",
        "name": "Alex Liss",
        "organization": "Huge"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Huge VP Alex Liss argues that awkward generative-AI additions undermine user trust and proposes applying AI to interface design rather than simply inserting chatbots into products. His workflow combines audience data from Huge's Live platform, intent mapping, simulated Intelligent Twins, computer-use and computer-vision models, and human oversight to accelerate needfinding and usability evaluation. A global sports-website audit illustrates simulated task completion and friction discovery, while MCP, Figma, React, and Node.js suggest future design-to-code integrations."
  },
  {
    "slug": "ionic-launch-opening-the-economy-to-ai-agents",
    "title": "Ionic Launch: Opening the economy to AI agents",
    "url": "https://ai.engineer/talks/ionic-launch-opening-the-economy-to-ai-agents",
    "videoId": "Ju9PeKEKb24",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 330000,
    "speakers": [
      {
        "slug": "justin-mccarty",
        "name": "Justin McCarty",
        "organization": "Ionic Commerce"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "finance",
        "name": "Finance"
      }
    ],
    "summary": "Ionic Commerce CEO and cofounder Justin McCarty presents infrastructure enabling AI agents to discover products, make personalized purchasing decisions, and complete transactions through a unified commerce API and payment tokens. He argues that merchant-funded monetization can outperform recurring consumer subscriptions and describes an existing merchant network and catalog containing millions of SKUs."
  },
  {
    "slug": "it-admin-for-the-ai-workforce-sarthak-aggarwal-decawork",
    "title": "IT Admin for the AI Workforce — Sarthak Aggarwal, Decawork",
    "url": "https://ai.engineer/talks/it-admin-for-the-ai-workforce-sarthak-aggarwal-decawork",
    "videoId": "q-WOjZhOMCA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 977000,
    "speakers": [
      {
        "slug": "sarthak-aggarwal",
        "name": "Sarthak Aggarwal",
        "organization": "Decawork"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Decawork co-founder Sarthak Aggarwal argues that enterprises must manage AI agents as a second workforce, providing runtime identities, accountable ownership, delegated permissions, auditing, and revocation. He examines OAuth delegation and emerging governance products, then contrasts the [REDACTED:username] vulnerability in Microsoft 365 Copilot with a Replit production-access incident to show why instructions and model guardrails are insufficient security boundaries. His proposed architecture separates trusted planning from untrusted content processing and constrains execution through explicit capabilities and enforceable policy."
  },
  {
    "slug": "it-s-10pm-do-you-know-where-your-agents-are",
    "title": "It's 10pm. Do You Know Where Your Agents Are?",
    "url": "https://ai.engineer/talks/it-s-10pm-do-you-know-where-your-agents-are",
    "videoId": "I3znWC3MEXM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1382000,
    "speakers": [
      {
        "slug": "kim-maida",
        "name": "Kim Maida",
        "organization": "Keycard"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Keycard's Kim Maida demonstrates how overprivileged incident-response agents can delete a production Postgres database, restart infrastructure, or incur cloud spending using broadly scoped API keys. She maps the agent–LLM–MCP execution path and explains how OAuth token exchange and a Security Token Service can apply identity and policy checks before issuing narrowly scoped, audience-restricted credentials for individual tool calls. The talk concludes with questions about MCP enforcement boundaries and adoption of existing OAuth standards."
  },
  {
    "slug": "iterating-on-llm-apps-at-scale-learnings-from-discord",
    "title": "Iterating on LLM apps at scale: Learnings from Discord",
    "url": "https://ai.engineer/talks/iterating-on-llm-apps-at-scale-learnings-from-discord",
    "videoId": "OrtBEBLMXdM",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1106000,
    "speakers": [
      {
        "slug": "ian-webster",
        "name": "Ian Webster",
        "organization": "Discord"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Discord engineer Ian Webster explains how teams evaluated and iterated on Clyde, a large-scale conversational AI product, while reducing safety and reliability risks. He advocates developer-first evaluation with Promptfoo, simple deterministic metrics, shorter model-specific prompts, occasional GPT-4 responses to steer conversations, and red teaming against application-specific jailbreaks."
  },
  {
    "slug": "judge-the-judge-building-llm-evaluators-that-actually-work-with-gepa-mahmoud-mabrouk-agenta-ai",
    "title": "Judge the Judge: Building LLM Evaluators That Actually Work with GEPA — Mahmoud Mabrouk, Agenta AI",
    "url": "https://ai.engineer/talks/judge-the-judge-building-llm-evaluators-that-actually-work-with-gepa-mahmoud-mabrouk-agenta-ai",
    "videoId": "X4dEHRzBLmc",
    "event": "AI Engineer Europe 2026",
    "durationMs": 2451000,
    "speakers": [
      {
        "slug": "mahmoud-mabrouk",
        "name": "Mahmoud Mabrouk",
        "organization": "Agenta"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Agenta co-founder Mahmoud Mabrouk presents a practical workshop on building LLM-as-a-Judge evaluators calibrated against human annotations. Using Sierra's τ-bench airline customer-support scenario and 599 annotated conversation traces, he explains business-specific metric design, four targeted judges, GEPA-based reflective prompt optimization, validation, and a production data flywheel linking observability with continuous evaluation improvements."
  },
  {
    "slug": "judging-llms",
    "title": "Judging LLMs",
    "url": "https://ai.engineer/talks/judging-llms",
    "videoId": "IIL2tE4n1Q0",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1119000,
    "speakers": [
      {
        "slug": "alex-volkov",
        "name": "Alex Volkov",
        "organization": "Weights & Biases"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Alex Volkov of Weights & Biases uses a comedic courtroom format to explain why production LLM applications need human oversight, comprehensive tracing, prompt and code versioning, and continuous evaluations. He demonstrates Weave for inspecting RAG and agent call hierarchies, comparing development and production examples, defining evaluation criteria, detecting bias, and visualizing experiments, including Chris Van Pelt’s OpenUI project."
  },
  {
    "slug": "just-do-it-let-your-tools-think-for-themselves-robert-chandler",
    "title": "Just do it. (let your tools think for themselves) - Robert Chandler",
    "url": "https://ai.engineer/talks/just-do-it-let-your-tools-think-for-themselves-robert-chandler",
    "videoId": "lp0pswT_FEI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 410000,
    "speakers": [
      {
        "slug": "robert-chandler",
        "name": "Robert Chandler",
        "organization": "Wordware"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Wordware co-founder and CTO Robert Chandler argues that thin MCP wrappers around conventional APIs produce unreliable agents by exposing excessive tools, verbose JSON, pagination, authentication friction, and long reasoning chains. After describing a failed Slack messaging task, he proposes specialized agentic tools with natural-language interfaces and demonstrates Wordware MCP Toolbox orchestrating Claude, Twitter data, and Notion to create an Anthropic competitor analysis."
  },
  {
    "slug": "keynote-the-ai-developer-experience-doesn-t-have-to-suck-why-and-how-we-built-modal",
    "title": "Keynote: The AI developer experience doesn't have to suck – why and how we built Modal",
    "url": "https://ai.engineer/talks/keynote-the-ai-developer-experience-doesn-t-have-to-suck-why-and-how-we-built-modal",
    "videoId": "qeDPKbWjsuk",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1298000,
    "speakers": [],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Modal founder and CEO Erik [REDACTED:username] explains how the company built managed, Python-first cloud infrastructure to improve the developer experience for AI and other compute-intensive applications. He contrasts Modal with Kubernetes, AWS Lambda, and model APIs; discusses generative AI inference, Suno, diffusion, computational biology, and large-scale batch processing; and demonstrates serverless functions, NVIDIA GPUs, Python-defined container images, autoscaling, operational metrics, image caching, and lazy loading."
  },
  {
    "slug": "keynote-why-people-think-agent-is-a-buzzword-but-it-isn-t",
    "title": "Keynote: Why people think \"agent\" is a buzzword but it isn't",
    "url": "https://ai.engineer/talks/keynote-why-people-think-agent-is-a-buzzword-but-it-isn-t",
    "videoId": "D6v5rlqUIc8",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1687000,
    "speakers": [
      {
        "slug": "chip-huyen",
        "name": "Chip Huyen"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Chip Huyen explains why AI agents represent a substantive engineering paradigm: models perceive environments, use tools, and plan multistep actions across APIs, coding systems, and workplace software. She examines how task complexity increases failure rates, compares reasoning and non-reasoning models on planning tasks, and uses ZebraLogic and Z3 conflicts to illustrate scaling limits. She also discusses ambiguous instructions, specialized action models, growing tool context, and memory systems."
  },
  {
    "slug": "knowledge-graphs-graphrag-techniques-for-building-effective-genai-applications",
    "title": "Knowledge Graphs & GraphRAG: Techniques for Building Effective GenAI Applications",
    "url": "https://ai.engineer/talks/knowledge-graphs-graphrag-techniques-for-building-effective-genai-applications",
    "videoId": "2DyHW23L6Cs",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 5992000,
    "speakers": [
      {
        "slug": "andreas-kolleger",
        "name": "Andreas Kolleger",
        "organization": "[REDACTED:username]"
      },
      {
        "slug": "zach-blumenthal",
        "name": "Zach Blumenthal"
      },
      {
        "slug": "michael-hunger",
        "name": "Michael Hunger"
      },
      {
        "slug": "tomasz",
        "name": "Tomasz"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "A hands-on [REDACTED:username] GraphRAG workshop led by a presenter introducing himself as Zach, with Andreas Kolleger and other helpers present, walks through a Python notebook and [REDACTED:username] Graph Data Science Sandbox. Participants configure a Bolt connection and GPT-4o, query graph data with Cypher, create embeddings, examine Lucene-backed HNSW vector search and LangChain integration, and apply collaborative filtering, node embeddings, and graph projections to personalized recommendations and generated responses."
  },
  {
    "slug": "knowledge-graphs-in-litigation-agents-tom-smoker-whyhow-ai",
    "title": "Knowledge Graphs in Litigation Agents — Tom Smoker, WhyHow.AI",
    "url": "https://ai.engineer/talks/knowledge-graphs-in-litigation-agents-tom-smoker-whyhow-ai",
    "videoId": "yYxr6LdXNWM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1153000,
    "speakers": [
      {
        "slug": "tom-smoker",
        "name": "Tom Smoker",
        "organization": "WhyHow.AI"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "WhyHow.AI technical founder Tom Smoker explains how knowledge graphs, GraphRAG, and multi-agent systems turn documents and public web complaints into structured, personalized litigation intelligence for lawyers. He describes legal-domain schemas, discovery signals, the compounding reliability costs of chained LLM workflows, and iterative graph-state management that improves reporting and case identification."
  },
  {
    "slug": "latent-space-paper-club-aiewf-special-edition-test-of-time-deepseek-r1-v3-vibhu-sapra",
    "title": "Latent Space Paper Club: AIEWF Special Edition (Test of Time, DeepSeek-R1/V3) — Vibhu Sapra",
    "url": "https://ai.engineer/talks/latent-space-paper-club-aiewf-special-edition-test-of-time-deepseek-r1-v3-vibhu-sapra",
    "videoId": "9k3xPh-40mo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 3234000,
    "speakers": [
      {
        "slug": "vibhu-sapra",
        "name": "Vibhu Sapra",
        "organization": "Latent.Space"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Vibhu Sapra reviews Latent Space Paper Club’s community growth and launches Test of Time Paper Club, a curriculum-oriented series on foundational AI papers and engineering concepts. He then examines DeepSeek-V3 and DeepSeek-R1, discussing reasoning benchmarks, test-time compute, mixture-of-experts models, knowledge distillation, reinforcement learning, GRPO versus PPO, rejection sampling and the broader adoption of open reasoning models."
  },
  {
    "slug": "lawrence-jones-fighting-ai-with-ai",
    "title": "Lawrence Jones - Fighting AI with AI",
    "url": "https://ai.engineer/talks/lawrence-jones-fighting-ai-with-ai",
    "videoId": "L2r6vLlLgs8",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1049000,
    "speakers": [
      {
        "slug": "lawrence-jones",
        "name": "Lawrence Jones",
        "organization": "incident.io"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "incident.io founding engineer Lawrence Jones explains how the team uses coding agents to debug its complex AI SRE systems. He describes treating prompt evaluations as Go-adjacent YAML unit tests, building an agent-friendly CLI to manage oversized evaluation suites, exporting debugging interfaces and prompt hierarchies as filesystems for Claude Code, and using repeatable analysis, backtests, and red-green evaluation loops to validate fixes."
  },
  {
    "slug": "layering-every-technique-in-rag-one-query-at-a-time",
    "title": "Layering every technique in RAG, one query at a time",
    "url": "https://ai.engineer/talks/layering-every-technique-in-rag-one-query-at-a-time",
    "videoId": "w9u11ioHGA0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1222000,
    "speakers": [
      {
        "slug": "david-karam",
        "name": "David Karam",
        "organization": "Pi Labs"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "David Karam of Pi Labs, formerly of Google Search, explains how to improve retrieval-augmented generation incrementally through a quality-engineering loop grounded in real queries and observed failures. He compares BM25, vector retrieval, custom embeddings, and cross-encoder reranking; examines cost and relevance limitations; introduces domain-specific ranking signals such as price and popularity; and describes query orchestration and multi-backend retrieval for ambiguous search intent."
  },
  {
    "slug": "leadership-in-ai-assisted-engineering",
    "title": "Leadership in AI-Assisted Engineering",
    "url": "https://ai.engineer/talks/leadership-in-ai-assisted-engineering",
    "videoId": "PmZDupFP3UM",
    "event": "AI Engineer Code 2025",
    "durationMs": 1091000,
    "speakers": [
      {
        "slug": "justin-reock",
        "name": "Justin Reock",
        "organization": "DX"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Justin Reock of DX explains how engineering leaders can move beyond AI adoption mandates and perceived productivity toward evidence-based measurement of software delivery, code quality, and business impact. Drawing on METR, DORA, DX data, Project Aristotle, and SWE-bench, he recommends clear policies, experimentation time, psychological safety, transparent metrics, and applying AI across SDLC bottlenecks rather than focusing narrowly on code generation."
  },
  {
    "slug": "learned-execution-graphs-for-anomaly-detection-drift-in-apis-ritvik-pandya-jp-morgan-chase",
    "title": "Learned Execution Graphs for Anomaly Detection & Drift in APIs — Ritvik Pandya, JP Morgan Chase",
    "url": "https://ai.engineer/talks/learned-execution-graphs-for-anomaly-detection-drift-in-apis-ritvik-pandya-jp-morgan-chase",
    "videoId": "u1yaOeEX4e8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1178000,
    "speakers": [
      {
        "slug": "ritvik-pandya",
        "name": "Ritvik Pandya",
        "organization": "JP Morgan Chase"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "finance",
        "name": "Finance"
      }
    ],
    "summary": "Ritvik Pandya describes representing payment API requests as short-lived, learned execution DAGs to identify anomalous processing paths, localize failures to individual nodes, and distinguish isolated anomalies from longer-term drift. The approach combines client- and payment-specific baselines, divergence-based analysis, OpenTelemetry traces, asynchronous Kafka processing, sampling, and explainable operational decisions."
  },
  {
    "slug": "learning-on-the-job-the-future-of-post-training",
    "title": "Learning on the job: the future of post-training",
    "url": "https://ai.engineer/talks/learning-on-the-job-the-future-of-post-training",
    "videoId": "k35LeKZEhiE",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1100000,
    "speakers": [
      {
        "slug": "raymond-feng",
        "name": "Raymond Feng",
        "organization": "Applied Compute"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Applied Compute’s Raymond Feng explains how agent post-training can progress from single-turn questions and synthetic environments toward long-horizon enterprise tasks and continual learning from real interactions. He describes replayable orchestrator-and-sandbox environments, GRPO’s comparison of multiple rollouts, the difficulty of reproducing production conditions, and NVIDIA’s Polar framework for reinforcement learning across existing agent harnesses."
  },
  {
    "slug": "lessons-from-building-genai-based-applications-juan-peredo",
    "title": "Lessons from building GenAI based applications — Juan Peredo",
    "url": "https://ai.engineer/talks/lessons-from-building-genai-based-applications-juan-peredo",
    "videoId": "YYcNm2RexnY",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1993000,
    "speakers": [
      {
        "slug": "juan-peredo",
        "name": "Juan Peredo"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Juan Peredo presents practical lessons from building generative-AI applications, contrasting coding-assistant hype with the operational complexity of production AI. He discusses selecting models and developer tools, mitigating hallucinations through RAG and output validation, choosing local or cloud hosting, controlling infrastructure and agent tool-calling costs, and using evaluation and observability to monitor changing models and diagnose failures."
  },
  {
    "slug": "lessons-from-building-linkedin-s-genai-platform",
    "title": "Lessons from Building LinkedIn's GenAI Platform",
    "url": "https://ai.engineer/talks/lessons-from-building-linkedin-s-genai-platform",
    "videoId": "n9rjuBuShko",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1073000,
    "speakers": [
      {
        "slug": "xiaofeng-wang",
        "name": "Xiaofeng Wang",
        "organization": "LinkedIn"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "LinkedIn's Xiaofeng Wang describes the evolution of its generative AI platform from GPT-4-powered Collaborative Articles to retrieval-augmented coaching experiences and the multi-agent Hiring Assistant. The platform combines a LangChain-based Python SDK, integration with LinkedIn's Java infrastructure, prompt versioning, conversational memory, distributed agent orchestration, a centralized API skill registry, and OTEL-based observability. Wang also discusses interchangeable model backends, responsible AI governance, and hiring a team with complementary technical skills."
  },
  {
    "slug": "lessons-from-studying-every-memory-system",
    "title": "Lessons from Studying Every Memory System",
    "url": "https://ai.engineer/talks/lessons-from-studying-every-memory-system",
    "videoId": "5ZGyKWjQDr0",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1171000,
    "speakers": [
      {
        "slug": "shlok-khemani",
        "name": "Shlok Khemani"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Independent researcher Shlok Khemani examines how consumer AI products implement personalized memory, tracing ChatGPT from explicit saved facts introduced in 2024 to broader conversation-informed memory in 2025. He discusses dense running profiles, background updates, stale or incorrectly inferred memories, limited profile visibility, and approaches to richer learning loops outside model weights."
  },
  {
    "slug": "let-llms-wander-engineering-rl-environments-stefano-fiorucci",
    "title": "Let LLMs Wander: Engineering RL Environments — Stefano Fiorucci",
    "url": "https://ai.engineer/talks/let-llms-wander-engineering-rl-environments-stefano-fiorucci",
    "videoId": "71V3fTaUp2Q",
    "event": "AI Engineer Europe 2026",
    "durationMs": 2435000,
    "speakers": [
      {
        "slug": "stefano-fiorucci",
        "name": "Stefano Fiorucci",
        "organization": "deepset"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Stefano Fiorucci, an AI/software engineer at deepset and Haystack contributor, explains how reinforcement-learning environments let language models improve through interaction, rewards, and multi-step tasks. He connects reasoning-model advances such as OpenAI o1, DeepSeek-R1, GRPO, and PPO to practical environment engineering with Prime Intellect's Verifiers and prime-rl. Through multi-turn examples and a tic-tac-toe environment with minimax-based opponents, he examines model serving, reward design, opponent difficulty, rollout inspection, batch-size-related instability, sampling temperature, and specialized small-model training with tools."
  },
  {
    "slug": "let-s-build-an-agent-from-scratch-kam-lasater",
    "title": "Let's Build an Agent from Scratch — Kam Lasater",
    "url": "https://ai.engineer/talks/let-s-build-an-agent-from-scratch-kam-lasater",
    "videoId": "xzXdLRUyjUg",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1169000,
    "speakers": [
      {
        "slug": "kam-lasater",
        "name": "Kam Lasater",
        "organization": "Manufactured"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Kam Lasater builds a minimal AI agent incrementally without relying on an agent framework. Starting with an OpenAI LLM call and a strict JSON LLM-as-judge, he adds SerpApi-backed Google Search, application-managed tool execution, iterative control flow, and a to-do-list-based planning and memory system to show how simple components become an agent."
  },
  {
    "slug": "let-s-go-bananas-with-genmedia",
    "title": "Let's go Bananas with GenMedia",
    "url": "https://ai.engineer/talks/let-s-go-bananas-with-genmedia",
    "videoId": "BcWFc3H7Khg",
    "event": "AI Engineer Europe 2026",
    "durationMs": 4634000,
    "speakers": [
      {
        "slug": "guillaume-vernade",
        "name": "Guillaume Vernade",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Google DeepMind developer advocate Guillaume Vernade leads a hands-on generative-media workshop covering image-model APIs, story-to-dialogue adaptation, synthetic character voices, and controllable music generation. Demonstrations explore notebook-based workflows, character accents, mixed musical genres, and song structure. An organizer subsequently explains that building-access problems affected attendance and announces an afternoon reprise."
  },
  {
    "slug": "let-s-integrate-ai-agents-in-event-sourced-systems",
    "title": "Let's integrate AI Agents in Event-Sourced Systems",
    "url": "https://ai.engineer/talks/let-s-integrate-ai-agents-in-event-sourced-systems",
    "videoId": "o6U_2vd967Y",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1297000,
    "speakers": [
      {
        "slug": "divakar-kumar",
        "name": "Divakar Kumar",
        "organization": "FlyersSoft"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Divakar Kumar of FlyersSoft explains how to augment existing event-sourced financial systems with AI agents for real-time fraud detection. A two-tier architecture keeps rule-based and machine-learning systems responsible for routine decisions while escalating ambiguous transactions to agentic orchestration. The approach assembles account, KYC, transaction, and device context through domain-driven bounded contexts, projections, semantic layers, and materialized views, using Azure Cosmos DB as an event store."
  },
  {
    "slug": "let-s-talk-about-fomat-fear-of-missing-agent-time",
    "title": "Let’s Talk About FOMAT – Fear of Missing Agent Time",
    "url": "https://ai.engineer/talks/let-s-talk-about-fomat-fear-of-missing-agent-time",
    "videoId": "W-SX_srBa3Y",
    "event": "AI Engineer Europe 2026",
    "durationMs": 977000,
    "speakers": [
      {
        "slug": "michael-richman",
        "name": "Michael Richman",
        "organization": "Bitly"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Bitly engineering leader Michael Richman introduces FOMAT, or fear of missing agent time, and demonstrates Cmd+Ctrl, his system for monitoring, interacting with, and launching coding-agent sessions remotely. Examples span Claude Code, Codex, Gemini, and Cursor, with mobile and web access, notifications when agents need attention, and centralized session organization."
  },
  {
    "slug": "letting-ai-interface-with-your-app-with-mcp",
    "title": "Letting AI Interface with Your App with MCP",
    "url": "https://ai.engineer/talks/letting-ai-interface-with-your-app-with-mcp",
    "videoId": "EyZiAp0pelw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1194000,
    "speakers": [
      {
        "slug": "kent-c-dodds",
        "name": "Kent C. Dodds",
        "organization": "EpicAI.pro"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Kent C. Dodds explains how Model Context Protocol lets AI assistants interact with applications, external tools, and services through a standardized interface. Using a Jarvis film clip, he illustrates multimodal interaction, dynamically generated interfaces, cross-system research, and the difficulty of supporting long-tail integrations. He argues that effective host applications and tool calling transform language models into useful assistants, then demonstrates authentication against an Epic Me server and points viewers to EpicAI.pro."
  },
  {
    "slug": "level-up-your-automation-game-agentic-workflows-on-vertex-ai",
    "title": "Level Up Your Automation Game: Agentic Workflows on Vertex AI",
    "url": "https://ai.engineer/talks/level-up-your-automation-game-agentic-workflows-on-vertex-ai",
    "videoId": "n0qemluQDtQ",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1086000,
    "speakers": [
      {
        "slug": "rukma-sen",
        "name": "Rukma Sen",
        "organization": "Google"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Google’s Rukma Sen presents AI agents as practical interfaces to generative AI and describes their core architecture: a reasoning model, action-taking tools and APIs, and orchestration that maintains goals, memory and state. She contrasts deterministic, generative and hybrid agents, illustrates multi-agent customer-service routing and supervision, surveys enterprise applications, and introduces Google Cloud’s Vertex AI as a platform for building agents and managing broader AI development."
  },
  {
    "slug": "llama-3-at-1-redacted-password-000-tok-s-on-the-sambanova-ai-platform",
    "title": "Llama 3 at 1000 tok/s on the SambaNova AI Platform",
    "url": "https://ai.engineer/talks/llama-3-at-1-redacted-password-000-tok-s-on-the-sambanova-ai-platform",
    "videoId": "z_Xh2MzqKDM",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 3658000,
    "speakers": [
      {
        "slug": "rachelle-mattern",
        "name": "Rachelle Mattern",
        "organization": "SambaNova Systems"
      },
      {
        "slug": "petro-milan",
        "name": "Petro Milan",
        "organization": "SambaNova"
      },
      {
        "slug": "varun-krishna",
        "name": "Varun Krishna",
        "organization": "SambaNova"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "SambaNova presenters introduce high-throughput Llama 3 inference and the Samba-1 Composition of Experts architecture[REDACTED:password] then demonstrate Llama 3 and Samba-1 Turbo before leading hands-on exercises. The workshop covers LangChain prompt composition[REDACTED:password] Python environment setup[REDACTED:password] SambaStudio inference and embedding APIs[REDACTED:password] document loading and splitting[REDACTED:password] ChromaDB indexing[REDACTED:password] and PDF-based enterprise retrieval-augmented generation."
  },
  {
    "slug": "llamafile-bringing-ai-to-the-masses-with-fast-cpu-inference",
    "title": "Llamafile: bringing AI to the masses with fast CPU inference",
    "url": "https://ai.engineer/talks/llamafile-bringing-ai-to-the-masses-with-fast-cpu-inference",
    "videoId": "-mRi-B3t6fA",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1045000,
    "speakers": [
      {
        "slug": "stephen-hood",
        "name": "Stephen Hood",
        "organization": "Mozilla"
      },
      {
        "slug": "justine-tunney",
        "name": "Justine Tunney",
        "organization": "Mozilla"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Stephen Hood and Justine Tunney introduce Mozilla-backed llamafile, an open-source project that packages large language models as portable, single-file executables for local use. They explain why faster consumer-CPU inference matters, how the project builds on and contributes to llama.cpp, and how matrix-multiplication and quantization improvements accelerate workloads across x86 and ARM. The presentation also covers offline operation, a summarization demonstration, and Mozilla Builders' support for open-source AI."
  },
  {
    "slug": "llm-codegen-fails-and-how-to-stop-em",
    "title": "LLM codegen fails and how to stop 'em",
    "url": "https://ai.engineer/talks/llm-codegen-fails-and-how-to-stop-em",
    "videoId": "juoNbJiZUi0",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1158000,
    "speakers": [
      {
        "slug": "danilo-campos",
        "name": "Danilo Campos",
        "organization": "PostHog"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Danilo Campos explains how PostHog Wizard reduces LLM code-generation failures when integrating PostHog into applications. His practical techniques include supplying current documentation to counter stale model knowledge, maintaining lightweight representative projects called Model Airplanes, and breadcrumbing agents through controlled implementation steps to reduce hallucinated APIs, architectural mistakes, and support burdens. He concludes that strong prose instructions retain value as models improve and that agents need useful boundaries without excessive constraint; the recording includes an audience question."
  },
  {
    "slug": "llm-evals-that-work-irl",
    "title": "LLM Evals That Work IRL",
    "url": "https://ai.engineer/talks/llm-evals-that-work-irl",
    "videoId": "nbZzSC5A6hs",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1129000,
    "speakers": [
      {
        "slug": "aparna-dhinakaran",
        "name": "Aparna Dhinakaran",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Arize AI co-founder Aparna Dhinakaran distinguishes general model benchmarks from application-specific LLM evaluations and explains LLM-as-a-judge, user feedback, and heuristic approaches. Using a commerce chatbot and Phoenix demo, she shows how to trace routing and function-calling failures, evaluate individual application components, construct datasets, and iterate through experiments. She concludes with practical guidance on explanatory versus numerical evaluations and research into needle-in-a-haystack testing, RAG context placement, and differences among foundation models."
  },
  {
    "slug": "llm-knowledge-bases-a-practical-guide",
    "title": "LLM Knowledge Bases: a practical guide",
    "url": "https://ai.engineer/talks/llm-knowledge-bases-a-practical-guide",
    "videoId": "I3bpdgFJCUY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1277000,
    "speakers": [
      {
        "slug": "ben-holmes",
        "name": "Ben Holmes",
        "organization": "Warp"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Warp developer-relations lead Ben Holmes demonstrates how to turn scattered Markdown notes into an LLM-accessible personal knowledge base. Using his Hubble app, Obsidian, Warp, and an agent skill called enrich note, he shows how agents can add sources, tags, and backlinks; generate interconnected wikis and visualizations; and automate updates using Obsidian CLI and Docker-backed cloud sandboxes."
  },
  {
    "slug": "llm-observability-evaluation-experimentation-platform-dat-ngo-arize-ai",
    "title": "LLM Observability, Evaluation, Experimentation Platform — Dat Ngo, Arize AI",
    "url": "https://ai.engineer/talks/llm-observability-evaluation-experimentation-platform-dat-ngo-arize-ai",
    "videoId": "JsCCrBF7F1g",
    "event": "AI Engineer Europe 2026",
    "durationMs": 992000,
    "speakers": [
      {
        "slug": "dat-ngo",
        "name": "Dat Ngo",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Arize AI's Dat Ngo explains how OpenTelemetry traces and spans expose nondeterministic agent behavior, regressions, and incorrectly ordered tool calls. He discusses Arize AX, LLM-as-a-judge evaluation calibrated against trusted golden datasets, lower-cost deterministic checks such as JSON validation, collaboration between engineers and domain experts, and minimizing manual observability and evaluation overhead."
  },
  {
    "slug": "llm-quality-optimization-bootcamp",
    "title": "LLM Quality Optimization Bootcamp",
    "url": "https://ai.engineer/talks/llm-quality-optimization-bootcamp",
    "videoId": "2Wtq2GvUicw",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 3185000,
    "speakers": [
      {
        "slug": "thierry-moreau",
        "name": "Thierry Moreau",
        "organization": "OctoAI"
      },
      {
        "slug": "pedro-torruella",
        "name": "Pedro Torruella",
        "organization": "OctoAI"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "This hands-on OctoAI workshop explains how to fine-tune smaller Llama 3 models for specialized tasks, including function calling, classification, and PII redaction. It contrasts deployment costs and GPU constraints with larger proprietary models, demonstrates OpenPipe fine-tuning and model-weight export, and walks through deploying checkpoints and LoRA assets with the OctoAI CLI. The presenter reports a specific 200-fold cost comparison with GPT-4 Turbo and notes a LoRA upload issue during the demonstration; neither observation independently establishes general performance or savings."
  },
  {
    "slug": "llm-safeguards-security-privacy-compliance-anti-hallucination",
    "title": "LLM Safeguards: Security, Privacy, Compliance, Anti-Hallucination",
    "url": "https://ai.engineer/talks/llm-safeguards-security-privacy-compliance-anti-hallucination",
    "videoId": "jdeMJJ_oNYg",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 2050000,
    "speakers": [
      {
        "slug": "daniel-whitenack",
        "name": "Daniel Whitenack",
        "organization": "Prediction Guard"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Prediction Guard founder and CEO Daniel Whitenack presents a practical checklist for deploying enterprise LLM applications safely, addressing hallucination risks in high-stakes settings, insecure open-model dependencies, sensitive-data and PII exposure, and prompt injection. Audience questions extend the discussion to firewall-style defenses, RAG data permissions, organizational data classification, and dry-run workflows with human review."
  },
  {
    "slug": "llm-scientific-reasoning-how-to-make-ai-capable-of-nobel-prize-discoveries",
    "title": "LLM Scientific Reasoning: How to Make AI Capable of Nobel Prize Discoveries",
    "url": "https://ai.engineer/talks/llm-scientific-reasoning-how-to-make-ai-capable-of-nobel-prize-discoveries",
    "videoId": "gbM7k3NauNI",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1200000,
    "speakers": [
      {
        "slug": "hubert-misztela",
        "name": "Hubert Misztela",
        "organization": "Novartis"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Novartis researcher Hubert Misztela examines whether large language models can accelerate scientific discovery by connecting findings across disconnected research domains. Using anomalous petunia gene-expression experiments as a motivating biological example, he critiques naive retrieval-augmented generation, distinguishes reasoning before and after retrieval, discusses GraphRAG and multi-concept retrieval, and outlines structured reasoning and retrospective discovery-evaluation approaches."
  },
  {
    "slug": "llms-for-the-working-programmer-become-a-10x-programming-centaur-today",
    "title": "LLMs for the working programmer. Become a 10x programming centaur today!",
    "url": "https://ai.engineer/talks/llms-for-the-working-programmer-become-a-10x-programming-centaur-today",
    "videoId": "zwItokY087U",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 4486000,
    "speakers": [
      {
        "slug": "manuel-odendahl",
        "name": "Manuel Odendahl",
        "organization": "The Tree Center"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Manuel Odendahl leads a hands-on workshop on using large language models to accelerate everyday software development. He introduces lightweight model and collaboration setup, discusses Mermaid diagrams and GitHub-based developer onboarding, and demonstrates domain-specific languages, simulated application interfaces, rapid prototyping, zero-shot prompting, and practical Copilot use."
  },
  {
    "slug": "lobster-trap-openclaw-in-containers-from-local-to-k8s-and-back",
    "title": "Lobster Trap: OpenClaw in Containers from Local to K8s and Back",
    "url": "https://ai.engineer/talks/lobster-trap-openclaw-in-containers-from-local-to-k8s-and-back",
    "videoId": "F1DYkY1BlfM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1316000,
    "speakers": [
      {
        "slug": "sally-ann-o-malley",
        "name": "Sally Ann O'Malley",
        "organization": "Red Hat"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Red Hat engineer Sally Ann O'Malley demonstrates running OpenClaw agents in portable containers that move between local Podman, Kubernetes, and OpenShift. She explains packaging agent tools, skills, and MCP servers; combining Podman secrets with OpenClaw secret references; preserving runtime state in volumes for backup and recovery; and launching specialized sub-agents across local and cluster environments."
  },
  {
    "slug": "local-agentic-theory-for-mobile-games-shafik-quoraishee-joanne-song-the-new-york-times",
    "title": "Local Agentic Theory For Mobile Games — Shafik Quoraishee & Joanne Song, The New York Times",
    "url": "https://ai.engineer/talks/local-agentic-theory-for-mobile-games-shafik-quoraishee-joanne-song-the-new-york-times",
    "videoId": "418t26CVz-w",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1084000,
    "speakers": [
      {
        "slug": "shafik-quoraishee",
        "name": "Shafik Quoraishee",
        "organization": "The New York Times"
      },
      {
        "slug": "joanne-song",
        "name": "Joanne Song",
        "organization": "The New York Times"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Shafik Quoraishee and Joanne Song present experimental on-device agents for accessible mobile games, emphasizing that existing New York Times puzzles are human-made and its games do not contain AI features. They contrast finite-state game logic and reinforcement-learning systems such as EfficientZero with agents that reason locally, demonstrate an agentic Space Invaders prototype, and discuss privacy, offline operation, rendering budgets, and 60 Hz frame constraints. The accessibility portion explores adapting gameplay to eye gaze, shaky taps, low vision, and handwriting, concluding with a vision of personalized AI running on individual devices."
  },
  {
    "slug": "local-models-trust-control-optimization",
    "title": "Local Models: Trust, Control, Optimization",
    "url": "https://ai.engineer/talks/local-models-trust-control-optimization",
    "videoId": "FWMJQDH3iK0",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2601000,
    "speakers": [
      {
        "slug": "carter-abdallah",
        "name": "Carter Abdallah",
        "organization": "NVIDIA"
      },
      {
        "slug": "vincent-weisser",
        "name": "Vincent Weisser",
        "organization": "Prime Intellect"
      },
      {
        "slug": "lucas-atkins",
        "name": "Lucas Atkins",
        "organization": "Arcee AI"
      },
      {
        "slug": "chris-alexiuk",
        "name": "Chris Alexiuk",
        "organization": "NVIDIA"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "A Local AI panel featuring publicly scheduled representatives of NVIDIA, Prime Intellect, and Arcee AI examines why builders want local and open models that provide ownership, durable access, privacy, and greater control than centralized APIs. Panelists describe collaboration around NVIDIA Nemotron and Arcee Trinity, discuss trust in open ecosystems and model provenance, and argue that efficient inference, accessible hardware, and developer participation are essential to advancing open frontier intelligence."
  },
  {
    "slug": "long-tasks-and-experienced-open-source-dev-productivity",
    "title": "Long Tasks and Experienced Open Source Dev Productivity",
    "url": "https://ai.engineer/talks/long-tasks-and-experienced-open-source-dev-productivity",
    "videoId": "k1t2xyWMUdY",
    "event": "AI Engineer Code 2025",
    "durationMs": 4552000,
    "speakers": [
      {
        "slug": "joel-becker",
        "name": "Joel Becker",
        "organization": "METR"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "METR researcher Joel Becker leads an interactive workshop on how AI compute growth and task-completion horizons relate to real-world software-development productivity. Participants examine agent-adoption J-curves, unreliable self-reported productivity, Cursor familiarity, and the challenges of evaluating experienced developers on natural tasks in mature open-source repositories. The discussion also considers AI Village as an example of agents attempting loosely specified, real-world goals."
  },
  {
    "slug": "loop-engineering-from-first-principles",
    "title": "Loop Engineering from first principles",
    "url": "https://ai.engineer/talks/loop-engineering-from-first-principles",
    "videoId": "xIt_mTQp6mY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1077000,
    "speakers": [
      {
        "slug": "kyle-mistele",
        "name": "Kyle Mistele",
        "organization": "HumanLayer"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "HumanLayer co-founder Kyle Mistele argues that autonomous coding agents should operate as measurable control systems rather than unbounded generation loops that produce enormous, unreviewable pull requests. He describes defining a desired codebase state, sensing violations with deterministic tools or agents, selecting bounded work, and actuating small changes through reviewable, scheduled workflows. A HumanLayer case study incrementally migrates an RPC API to Effect while tracking progress in version control and producing low-risk pull requests."
  },
  {
    "slug": "low-level-technicals-of-llms",
    "title": "Low Level Technicals of LLMs",
    "url": "https://ai.engineer/talks/low-level-technicals-of-llms",
    "videoId": "pRM_P6UfdIc",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 10346000,
    "speakers": [
      {
        "slug": "daniel-han",
        "name": "Daniel Han",
        "organization": "Unsloth"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Unsloth's Daniel Han leads an interactive workshop on low-level language-model internals and practical debugging. He examines Gemma, Phi-3, and Llama implementation problems; explains singular value decomposition, semantic embeddings, tokenization, token IDs, embedding lookup, and attention matrices; warns about training-data leakage and RoPE precision errors; and demonstrates fine-tuning workflows that save LoRA adapters."
  },
  {
    "slug": "luminal-search-based-deep-learning-compilers-joe-fioti",
    "title": "Luminal - Search-Based Deep Learning Compilers - Joe Fioti",
    "url": "https://ai.engineer/talks/luminal-search-based-deep-learning-compilers-joe-fioti",
    "videoId": "0uj9lMI-sIo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1475000,
    "speakers": [
      {
        "slug": "joe-fioti",
        "name": "Joe Fioti",
        "organization": "Luminal"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Luminal creator Joe Fioti explains how a small set of primitive tensor operations and static computation graphs can represent modern machine-learning models while avoiding the complexity of conventional framework stacks. He describes compiling those graphs directly to CUDA, searching candidate transformations to discover fast kernels instead of relying on handwritten optimization rules, handling bounded transformer dynamism, and exploiting whole-workload visibility for buffer reuse."
  },
  {
    "slug": "machines-of-buying-selling-grace",
    "title": "Machines of Buying & Selling Grace",
    "url": "https://ai.engineer/talks/machines-of-buying-selling-grace",
    "videoId": "zlZz0mDF2eg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1177000,
    "speakers": [
      {
        "slug": "adam-behrens",
        "name": "Adam Behrens",
        "organization": "New Generation (New Gen)"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "New Generation co-founder Adam Behrens examines how AI agents could transform online retail from static storefronts into conversational, programmatic commerce. He contrasts agent-optimized websites with merchant access through MCP servers and APIs, discusses dynamically generated shopping interfaces, and outlines checkout approaches involving virtual cards or Visa-style delegated authentication. He also addresses converting ambiguous customer intent into specific SKUs, integrating merchant product data, and the commercial potential of AI-referred shoppers."
  },
  {
    "slug": "magic-editor-under-the-hood-weaving-generative-ai-into-a-billion-user-app",
    "title": "Magic Editor Under the Hood: Weaving Generative AI into a Billion-User App",
    "url": "https://ai.engineer/talks/magic-editor-under-the-hood-weaving-generative-ai-into-a-billion-user-app",
    "videoId": "C13jiFWNuo8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1228000,
    "speakers": [
      {
        "slug": "kelvin-ma",
        "name": "Kelvin Ma",
        "organization": "Google Photos"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Google Photos engineer Kelvin Ma explains how computational photography and generative AI become production-ready mobile editing features. He describes a shared cross-platform C++ stack, on-device inference with TensorFlow Lite/LiteRT and Edge TPU, portrait segmentation, model-size and evaluation constraints, and Magic Eraser pipelines combining distractor detection, segmentation, inpainting, and graphics rendering. The talk emphasizes unpredictable edge cases and the reliability required to ship consumer AI features at scale."
  },
  {
    "slug": "make-your-llm-app-a-domain-expert-how-to-build-an-llm-native-expert-system",
    "title": "Make your LLM app a Domain Expert: How to Build an LLM-Native Expert System",
    "url": "https://ai.engineer/talks/make-your-llm-app-a-domain-expert-how-to-build-an-llm-native-expert-system",
    "videoId": "MRM7oA3JsFs",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1158000,
    "speakers": [
      {
        "slug": "christopher-lovejoy",
        "name": "Christopher Lovejoy",
        "organization": "Anterior"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Christopher Lovejoy, presenting for Anterior, explains how to build domain-native LLM applications around clinician expertise rather than model sophistication alone. Using healthcare authorization and Anterior’s Florence system, he describes an Adaptive Domain Intelligence Engine that categorizes failures in medical record extraction, clinical reasoning, and rules interpretation, incorporates clinician review, prioritizes harmful errors, and monitors regressions. He reports an improvement from approximately 95% to 99% accuracy and closes with audience questions about the expertise required."
  },
  {
    "slug": "make-your-own-event-sourced-agent-harness-using-stream-processors",
    "title": "Make your own event-sourced agent harness using stream processors",
    "url": "https://ai.engineer/talks/make-your-own-event-sourced-agent-harness-using-stream-processors",
    "videoId": "vi-2nasppAg",
    "event": "AI Engineer Europe 2026",
    "durationMs": 3867000,
    "speakers": [
      {
        "slug": "misha-kaletsky",
        "name": "Misha Kaletsky",
        "organization": "Iterate"
      },
      {
        "slug": "jonas-templestein",
        "name": "Jonas Templestein",
        "organization": "Iterate"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Iterate’s Jonas Templestein and Misha Kaletsky demonstrate an extensible event-sourced agent harness built around programmable event streams and stream processors. The workshop walks through a public GitHub repository, runnable cURL examples, a TypeScript SDK installed with pnpm, event-driven reducers and rate checks, JavaScript/TypeScript execution, and comparisons with cloud pub/sub offset management."
  },
  {
    "slug": "making-codebases-agent-ready",
    "title": "Making Codebases \"Agent-Ready\"",
    "url": "https://ai.engineer/talks/making-codebases-agent-ready",
    "videoId": "ShuJ_CN6zr4",
    "event": "AI Engineer Code 2025",
    "durationMs": 933000,
    "speakers": [
      {
        "slug": "eno-reyes",
        "name": "Eno Reyes",
        "organization": "Factory AI"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Factory's Eno Reyes argues that successful autonomous software engineering depends less on choosing marginally better coding models than on making codebases mechanically verifiable. He connects Software 2.0 and the asymmetry of verification to practical readiness measures: opinionated linters, unit and end-to-end tests, documented interfaces, OpenAPI specifications, and continuous validation. Using Factory's Droid as an example, he explains how specification-driven workflows and reliable pull-request checks enable parallel agents, modernization projects, and compounding feedback loops that improve engineering velocity."
  },
  {
    "slug": "making-open-models-10x-faster-and-better-for-modern-application-innovation",
    "title": "Making Open Models 10x faster and better for Modern Application Innovation",
    "url": "https://ai.engineer/talks/making-open-models-10x-faster-and-better-for-modern-application-innovation",
    "videoId": "C1CXwRYbwuQ",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1135000,
    "speakers": [
      {
        "slug": "lin-qiao",
        "name": "Lin Qiao",
        "organization": "Fireworks"
      },
      {
        "slug": "dmytro-dima-dzhulgakov",
        "name": "Dmytro (Dima) Dzhulgakov",
        "organization": "Fireworks AI"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Fireworks AI co-founder Dmytro Dzhulgakov substitutes for scheduled speaker Lin Qiao and explains how specialized open models can improve application latency, cost, and domain-specific quality. He discusses fine-tuned Llama and Gemma models, function calling, GPU deployment challenges, custom CUDA kernels, long-prompt RAG optimization and caching, multimodal image-model serving, and structured JSON generation."
  },
  {
    "slug": "malleable-evals-why-are-we-still-evaluating-adaptive-systems-with-static-tests",
    "title": "Malleable Evals: Why Are We Still Evaluating Adaptive Systems with Static Tests?",
    "url": "https://ai.engineer/talks/malleable-evals-why-are-we-still-evaluating-adaptive-systems-with-static-tests",
    "videoId": "4VhbYlfC7Gs",
    "event": "AI Engineer Europe 2026",
    "durationMs": 905000,
    "speakers": [
      {
        "slug": "vincent-koc",
        "name": "Vincent Koc",
        "organization": "Comet ML / OpenClaw"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Vincent Koc argues that fixed datasets, handcrafted test cases, and offline benchmarks cannot adequately evaluate AI agents whose behavior, tools, and operating conditions continually change. Drawing on OpenClaw, MCP-based agent components, and ARC-AGI reasoning benchmarks, he outlines malleable evaluations: trace-driven, always-on feedback systems that adapt alongside agents, optimize toward user intent, and help detect and correct emerging failures."
  },
  {
    "slug": "mastering-ai-pricing-mayank-pant-stripe",
    "title": "Mastering AI Pricing — Mayank Pant, Stripe",
    "url": "https://ai.engineer/talks/mastering-ai-pricing-mayank-pant-stripe",
    "videoId": "CrqPcIZOOXA",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1459000,
    "speakers": [
      {
        "slug": "mayank-pant",
        "name": "Mayank Pant",
        "organization": "Stripe"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Stripe billing solution architect Mayank Pant explains why variable AI compute costs undermine traditional SaaS pricing and presents hybrid subscription-plus-usage pricing as a way to protect margins while supporting adoption. Using Gamma and Intercom examples, he argues that billing units should reflect customer-perceived value and measurable outcomes rather than opaque API activity. He also discusses rate limits, churn, pricing experimentation, and audience questions about predictable costs and transaction-level usage tracking."
  },
  {
    "slug": "mastering-engineering-flow-with-windsurf",
    "title": "Mastering Engineering Flow with Windsurf",
    "url": "https://ai.engineer/talks/mastering-engineering-flow-with-windsurf",
    "videoId": "W_5tzQY-hVs",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 920000,
    "speakers": [
      {
        "slug": "eashan-sinha",
        "name": "Eashan Sinha",
        "organization": "Windsurf"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Windsurf engineer Eashan Sinha explains how the Windsurf Editor and its Cascade agent move beyond chat-based copilots toward collaborative, context-aware engineering workflows. He describes combining codebase retrieval, tool calling, observed developer actions, intent prediction, Windsurf Tab, MCP servers, and structured workflows, while arguing that developers should actively collaborate with coding agents rather than expect reliable unattended autonomy."
  },
  {
    "slug": "mastering-llm-inference-optimization-from-theory-to-cost-effective-deployment",
    "title": "Mastering LLM Inference Optimization: From Theory to Cost-Effective Deployment",
    "url": "https://ai.engineer/talks/mastering-llm-inference-optimization-from-theory-to-cost-effective-deployment",
    "videoId": "9tvJ_GYJA-o",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 2019000,
    "speakers": [
      {
        "slug": "mark-moyou",
        "name": "Mark Moyou",
        "organization": "NVIDIA"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "NVIDIA's Mark Moyou explains how to size and optimize production LLM inference deployments while controlling GPU costs. He connects tokenization, attention matrices, attention heads, and KV-cache behavior to throughput, input/output sequence lengths, and time-to-first-token measurements, then discusses NVIDIA inference tooling and adapting engines to changing traffic patterns during audience questions."
  },
  {
    "slug": "maximize-gpu-efficiency-with-continuous-profiling-for-gpus",
    "title": "Maximize GPU Efficiency with Continuous Profiling for GPUs",
    "url": "https://ai.engineer/talks/maximize-gpu-efficiency-with-continuous-profiling-for-gpus",
    "videoId": "wt8gzWR6auQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 691000,
    "speakers": [
      {
        "slug": "matthias-loibl",
        "name": "Matthias Loibl",
        "organization": "Polar Signals"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Matthias Loibl of Polar Signals explains continuous GPU profiling that combines NVIDIA NVML utilization, memory, power, temperature, clock-speed, and PCIe metrics with low-overhead Linux eBPF CPU profiles. He demonstrates correlating GPU underutilization with Python and CUDA call stacks in flame charts, measuring CUDA kernel execution time, and deploying the profiler through systemd, Docker, or Kubernetes."
  },
  {
    "slug": "mcp-agent-fine-tuning-workshop-ronan-mcgovern",
    "title": "MCP Agent Fine-Tuning Workshop - Ronan McGovern",
    "url": "https://ai.engineer/talks/mcp-agent-fine-tuning-workshop-ronan-mcgovern",
    "videoId": "Nqb7JTx0Pqo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2130000,
    "speakers": [
      {
        "slug": "ronan-mcgovern",
        "name": "Ronan McGovern",
        "organization": "Trelis Research"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Ronan McGovern of Trelis Research demonstrates how to connect a browser-capable agent to Model Context Protocol servers, expose a Qwen-compatible model through an OpenAI-style endpoint, and translate tool descriptions and Hermes-style JSON/XML calls. The workshop collects successful multi-turn reasoning and browser-use traces, prepares a small training dataset, and examines Qwen3 fine-tuning and tool-calling performance."
  },
  {
    "slug": "mcp-apps-extending-the-frontier-ido-salomon-liad-yosef",
    "title": "MCP Apps: Extending the Frontier — Ido Salomon & Liad Yosef",
    "url": "https://ai.engineer/talks/mcp-apps-extending-the-frontier-ido-salomon-liad-yosef",
    "videoId": "-jY2T2PiJBE",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1118000,
    "speakers": [
      {
        "slug": "ido-salomon",
        "name": "Ido Salomon",
        "organization": "MCP Apps"
      },
      {
        "slug": "liad-yosef",
        "name": "Liad Yosef",
        "organization": "MCP Apps"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "MCP Apps specification authors Ido Salomon and Liad Yosef explain how MCP-UI evolved into an open Model Context Protocol extension that lets services deliver branded, interactive application interfaces inside AI assistants instead of text-only responses. They outline host-to-application communication, ecosystem adoption, community governance, an interactive funnel example, and the need to avoid repeatedly rendering demanding interfaces such as 3D applications."
  },
  {
    "slug": "mcp-apps-primitives-discovery-and-the-future-of-software",
    "title": "MCP Apps: Primitives, Discovery, and the Future of Software",
    "url": "https://ai.engineer/talks/mcp-apps-primitives-discovery-and-the-future-of-software",
    "videoId": "sAOBXCDiDOs",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1734000,
    "speakers": [
      {
        "slug": "pietro-zullo",
        "name": "Pietro Zullo",
        "organization": "Manufact, Inc"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Manufact co-founder Pietro Zullo explains how MCP Apps extend Model Context Protocol servers with interactive, host-rendered interfaces, bidirectional communication, streamed tool inputs, and selectively shared model context. He introduces Manufact’s mcp-use SDKs, Inspector, and cloud platform; traces MCP-UI and cross-host adoption; discusses privacy-preserving application patterns; and examines distribution, discovery, and practical workflows using tools such as Granola and Linear."
  },
  {
    "slug": "mcp-is-all-you-need",
    "title": "MCP is all you need",
    "url": "https://ai.engineer/talks/mcp-is-all-you-need",
    "videoId": "bmWZk9vTze0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 924000,
    "speakers": [
      {
        "slug": "samuel-colvin",
        "name": "Samuel Colvin",
        "organization": "Pydantic"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Pydantic creator Samuel Colvin argues that Model Context Protocol can simplify communication among autonomous agents beyond its original desktop-assistant use cases. He contrasts MCP with OpenAPI, explains how sampling lets agentic MCP servers access an LLM through their client, and demonstrates Pydantic AI querying PyPI download data in BigQuery with retries and Logfire observability."
  },
  {
    "slug": "mcp-is-not-good-yet-david-cramer-sentry",
    "title": "MCP Is Not Good Yet — David Cramer, Sentry",
    "url": "https://ai.engineer/talks/mcp-is-not-good-yet-david-cramer-sentry",
    "videoId": "FCi4jT86gSw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1001000,
    "speakers": [
      {
        "slug": "david-cramer",
        "name": "David Cramer",
        "organization": "Sentry"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Sentry founder David Cramer offers a skeptical, practical assessment of Model Context Protocol, describing it as a pluggable architecture for agents rather than an inherently transformative technology. Using Sentry’s MCP server and integrations with Cursor and VS Code, he explains how monitoring and error context can reach coding environments, why remote transports and OAuth matter for B2B SaaS, and how excessive API calls and unconstrained responses can shift costs onto users."
  },
  {
    "slug": "mcp-mega-context-problem-matt-carey",
    "title": "MCP = Mega Context Problem - Matt Carey",
    "url": "https://ai.engineer/talks/mcp-mega-context-problem-matt-carey",
    "videoId": "YBYUvGOuotE",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1362000,
    "speakers": [
      {
        "slug": "matt-carey",
        "name": "Matt Carey",
        "organization": "Cloudflare"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Cloudflare's Matt Carey explains why exposing every API endpoint as an MCP tool overwhelms agent context windows: he describes an OpenAPI specification of roughly 2.3 million tokens, approximately 2,600 endpoints, and fragmented product-specific MCP servers. He contrasts that approach with self-discoverable CLIs, progressive tool discovery, and executing generated code, illustrating Wrangler and D1 before directing attendees to Cloudflare's agents package."
  },
  {
    "slug": "mcp-origins-and-requests-for-startups-theodora-chu-model-context-protocol-pm-anthropic",
    "title": "MCP: Origins and Requests For Startups — Theodora Chu, Model Context Protocol PM, Anthropic",
    "url": "https://ai.engineer/talks/mcp-origins-and-requests-for-startups-theodora-chu-model-context-protocol-pm-anthropic",
    "videoId": "x-8pBqWiTzk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1065000,
    "speakers": [
      {
        "slug": "theodora-chu",
        "name": "Theodora Chu",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Anthropic product manager Theodora Chu traces the Model Context Protocol from its creators’ frustration with manually copying external context into language models through internal development, its November 2024 open-source launch, and adoption by coding tools and major AI companies. She explains MCP’s emphasis on model agency, interoperable client-server integrations, and simple server-builder tooling, then frames the expanding ecosystem as an opportunity for startup founders."
  },
  {
    "slug": "mcp-tasks-async-why-the-heck-aren-t-any-agents-supporting-mcp-tasks-async",
    "title": "MCP Tasks (async)/ Why the heck aren't any agents supporting MCP tasks/async?",
    "url": "https://ai.engineer/talks/mcp-tasks-async-why-the-heck-aren-t-any-agents-supporting-mcp-tasks-async",
    "videoId": "s4r6nk5WsZw",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1434000,
    "speakers": [
      {
        "slug": "cornelia-davis",
        "name": "Cornelia Davis",
        "organization": "Temporal"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Temporal technologist Cornelia Davis explains why agent clients have been slow to adopt experimental MCP Tasks: durable, asynchronous tool execution introduces state management, failure recovery, client-side protocol complexity, and human-in-the-loop coordination. Using an ERP invoice-processing example and FastMCP demonstrations, she contrasts the stateful Tasks v1 design with a revised v2 approach and discusses resilient client/server architecture for long-running operations."
  },
  {
    "slug": "mcp-ui-extending-the-frontier-liad-yosef-and-ido-salomon-mcp-apps",
    "title": "MCP-UI: Extending the Frontier — Liad Yosef and Ido Salomon, MCP Apps",
    "url": "https://ai.engineer/talks/mcp-ui-extending-the-frontier-liad-yosef-and-ido-salomon-mcp-apps",
    "videoId": "o-zkvb0iFDQ",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1341000,
    "speakers": [
      {
        "slug": "liad-yosef",
        "name": "Liad Yosef",
        "organization": "Erel Labs"
      },
      {
        "slug": "ido-salomon",
        "name": "Ido Salomon"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "MCP Apps co-creators Liad Yosef and Ido Salomon explain how MCP-UI evolved into the first official Model Context Protocol extension for embedding branded, interactive application interfaces inside AI assistants. They contrast rich, context-preserving interfaces with text-only tool responses, describe bidirectional communication among applications, hosts, and backends, discuss adoption across assistants and developer tools, and argue that agent-hosted interfaces create a new application-distribution model."
  },
  {
    "slug": "mcps-are-boring-or-why-we-are-losing-the-sparkle-of-llms",
    "title": "MCPs are Boring (or: Why we are losing the Sparkle of LLMs)",
    "url": "https://ai.engineer/talks/mcps-are-boring-or-why-we-are-losing-the-sparkle-of-llms",
    "videoId": "J3oJqan2Gv8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1712000,
    "speakers": [
      {
        "slug": "manuel-odendahl",
        "name": "Manuel Odendahl"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Manuel Odendahl argues that conventional MCP integrations and rigid schema-defined tool calls underuse the expressive capabilities of large language models. Using weather and CRM examples, he illustrates tool proliferation, oversized responses, and brittle lookup behavior, then points to Voyager and demonstrates a JavaScript-generating, interactive CRM prototype as a more flexible approach to agent actions and open-ended creation."
  },
  {
    "slug": "measuring-agi-interactive-reasoning-benchmarks",
    "title": "Measuring AGI: Interactive Reasoning Benchmarks",
    "url": "https://ai.engineer/talks/measuring-agi-interactive-reasoning-benchmarks",
    "videoId": "3XmFPwjG8pg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1108000,
    "speakers": [
      {
        "slug": "greg-kamradt",
        "name": "Greg Kamradt",
        "organization": "ARC Prize Foundation"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "ARC Prize Foundation president Greg Kamradt argues that AGI benchmarks should measure how efficiently systems acquire unfamiliar skills relative to humans, rather than reward memorized game performance. He explains François Chollet's skill-acquisition framework and ARC-AGI grid tasks, then previews interactive, game-like ARC-AGI-3 evaluations requiring agents to explore, infer rules, and adapt. Audience questions address efficiency metrics and why tasks beyond human ability fall outside this benchmark's scope."
  },
  {
    "slug": "medic-for-apache-spark-first-aid-for-failing-jobs-drasko-profirovic-pinterest",
    "title": "Medic for Apache Spark - First Aid for Failing Jobs - Drasko Profirovic, Pinterest",
    "url": "https://ai.engineer/talks/medic-for-apache-spark-first-aid-for-failing-jobs-drasko-profirovic-pinterest",
    "videoId": "0RNNfxpdbQk",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 681000,
    "speakers": [
      {
        "slug": "drasko-profirovic",
        "name": "Drasko Profirovic",
        "organization": "Pinterest"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Pinterest staff engineer Drasko Profirovic explains Medic for Apache Spark, an agentic system for diagnosing failed Spark jobs. The architecture evolved from MCP-enabled tools and a single ReAct agent into an observable, regression-tested multi-agent harness built with LangGraph and Deep Agents. Key improvements include targeted exception retrieval and isolated multimodal analysis of time-series metrics, producing bounded-token visual summaries. The system also supports Spark SQL optimization, with potential expansion to Flink and Trino."
  },
  {
    "slug": "memory-harnesses-for-long-running-research-agents",
    "title": "Memory Harnesses for Long-Running Research Agents",
    "url": "https://ai.engineer/talks/memory-harnesses-for-long-running-research-agents",
    "videoId": "R3-anFK1YM8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 784000,
    "speakers": [
      {
        "slug": "stefania-druga",
        "name": "Stefania Druga",
        "organization": "Sakana AI"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Sakana AI research scientist Stefania Druga examines how context rot and forgotten decisions undermine long-running research agents, then presents an on-device memory harness combining persistent core traces, archival memory, vector retrieval, decision ledgers, and ranked recall. Using quantized Qwen 27B and DeepSeek V4 Flash on an M3 Ultra, she reports that ranked recall improves benchmark performance across XBench and Spider 2.0 while reducing wasted tokens, and discusses serial-inference constraints and sovereign AI."
  },
  {
    "slug": "memory-masterclass-make-your-ai-agents-remember-what-they-do-mark-bain-aius",
    "title": "Memory Masterclass: Make Your AI Agents Remember What They Do! — Mark Bain, AIUS",
    "url": "https://ai.engineer/talks/memory-masterclass-make-your-ai-agents-remember-what-they-do-mark-bain-aius",
    "videoId": "gsedOXz8FX4",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 3085000,
    "speakers": [
      {
        "slug": "mark-bain",
        "name": "Mark Bain",
        "organization": "AIUS Technologies"
      },
      {
        "slug": "vasilije-markovic",
        "name": "Vasilije Markovic",
        "organization": "Cognee"
      },
      {
        "slug": "daniel-chalef",
        "name": "Daniel Chalef",
        "organization": "Zep"
      },
      {
        "slug": "alex-gilmore",
        "name": "Alex Gilmore",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Mark Bain leads a workshop on durable AI-agent memory, arguing for causal relationships, knowledge graphs, and GraphRAG. Guest presenters Vasilije Markovic of Cognee, Alex Gilmore of Neo4j, and Daniel Chalef of Zep/Graphiti demonstrate graph-based agent workflows, memory-server integration with Claude Desktop, and temporal graphs. Bain compares MCP integrations across Neo4j, Graphiti, Cognee, and Mem0 before introducing a GraphRAG chat arena and discussing agentic firewalls and audience questions."
  },
  {
    "slug": "mentoring-the-machine",
    "title": "Mentoring the Machine",
    "url": "https://ai.engineer/talks/mentoring-the-machine",
    "videoId": "Zniw5c9_jx8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1146000,
    "speakers": [
      {
        "slug": "eric-hou",
        "name": "Eric Hou",
        "organization": "Augment Code"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Augment Code engineer Eric Hou argues that effective AI-assisted software development requires mentoring agents like junior engineers: provide outcomes, context, constraints, and architectural guidance rather than micromanaging implementation. Through a simulated overloaded workday, he demonstrates delegating RFC preparation, incident investigation, Git bisect, and a large gRPC remediation across parallel agents while engineers retain oversight. He concludes with recommendations for team-wide adoption and introduces Augment's Remote Agents."
  },
  {
    "slug": "mergeable-by-default-building-the-context-engine-to-save-time-and-tokens",
    "title": "Mergeable by default: Building the context engine to save time and tokens",
    "url": "https://ai.engineer/talks/mergeable-by-default-building-the-context-engine-to-save-time-and-tokens",
    "videoId": "5ID22ACI7IM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 6085000,
    "speakers": [
      {
        "slug": "peter-werry",
        "name": "Peter Werry",
        "organization": "Unblocked"
      },
      {
        "slug": "brandon",
        "name": "Brandon"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Unblocked founding engineer Peter Werry explains how context engines help coding agents incorporate organizational knowledge, team conventions, and prior engineering decisions without relying solely on naive retrieval or larger context windows. He discusses social engineering graphs, targeted and personalized retrieval, satisfaction-of-search failures, answer-caching pitfalls, ticket enrichment, and context-aware pull-request review. The recording includes audience Q&A and an extended hands-on workshop with an introduced colleague."
  },
  {
    "slug": "milliseconds-to-magic-real-time-workflows-using-the-gemini-live-api-and-pipecat",
    "title": "Milliseconds to Magic: Real‑Time Workflows using the Gemini Live API and Pipecat",
    "url": "https://ai.engineer/talks/milliseconds-to-magic-real-time-workflows-using-the-gemini-live-api-and-pipecat",
    "videoId": "fWY1FQwpWkY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1303000,
    "speakers": [
      {
        "slug": "kwindla-kramer",
        "name": "Kwindla Kramer",
        "organization": "Daily"
      },
      {
        "slug": "shrestha-basu-mallick",
        "name": "Shrestha Basu Mallick",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Kwindla Kramer of Daily and Shrestha Basu Mallick of Google DeepMind describe a layered real-time voice AI stack combining language models, Google's Gemini Live API, and Pipecat orchestration. They examine latency, application-versus-API responsibilities, and configurable server-side turn detection, then demonstrate a conversational agent managing tasks and on-screen lists. Examples connect voice interfaces to translation, education, healthcare, enterprise software, and more natural human-computer interaction."
  },
  {
    "slug": "mind-the-gap-in-your-agent-observability",
    "title": "Mind the Gap (In your Agent Observability)",
    "url": "https://ai.engineer/talks/mind-the-gap-in-your-agent-observability",
    "videoId": "iOXM3zE-2dk",
    "event": "AI Engineer Europe 2026",
    "durationMs": 4807000,
    "speakers": [
      {
        "slug": "amy-boyd",
        "name": "Amy Boyd",
        "organization": "Microsoft"
      },
      {
        "slug": "nitya-narasimhan",
        "name": "Nitya Narasimhan",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Microsoft presenters Amy Boyd and Nitya Narasimhan deliver a hands-on Microsoft Foundry workshop on closing observability gaps in AI agents. They introduce fleet monitoring and an evolving AIE Europe workshop repository, configure cloud resources and agent clients, demonstrate multi-turn interactions and Azure Monitor tracing, and assess quality, safety, groundedness, agentic behavior, and adversarial prompt attacks."
  },
  {
    "slug": "minimax-m2",
    "title": "Minimax M2",
    "url": "https://ai.engineer/talks/minimax-m2",
    "videoId": "lY1iFbDPRlw",
    "event": "AI Engineer Code 2025",
    "durationMs": 821000,
    "speakers": [
      {
        "slug": "olive-song",
        "name": "Olive Song",
        "organization": "MiniMax"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "MiniMax researcher Olive Song presents MiniMax-M2, an open-weight coding and agentic model with 10 billion active parameters. She explains developer-informed reinforcement learning and expert reward models, interleaved reasoning and tool use for long-horizon tasks, robustness across prompts and environments, and economical multi-agent workflows. The talk closes with prospective improvements in coding, memory, context management, and multimodal integration."
  },
  {
    "slug": "missing-pieces-of-workflow-automation",
    "title": "Missing pieces of workflow automation",
    "url": "https://ai.engineer/talks/missing-pieces-of-workflow-automation",
    "videoId": "xq9sz5MksVI",
    "event": "AI Engineer Summit 2025",
    "durationMs": 877000,
    "speakers": [
      {
        "slug": "shirsha-chaudhuri",
        "name": "Shirsha Chaudhuri",
        "organization": "Thomson Reuters"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Shirsha Chaudhuri describes Thomson Reuters’ progression from its Open Arena large-language-model playground through RAG and prompt engineering toward agents that automate complete enterprise workflows. Customer-support escalation and editorial-production examples illustrate opportunities for automation while retaining human approvals. She identifies integration connectors for existing enterprise systems, legacy mainframes, collaboration with domain specialists, and standardization as essential missing pieces for dependable workflow redesign."
  },
  {
    "slug": "mission-critical-evals-at-scale-learnings-from-100-000-medical-decisions",
    "title": "Mission-Critical Evals at Scale: Learnings from 100,000 Medical Decisions",
    "url": "https://ai.engineer/talks/mission-critical-evals-at-scale-learnings-from-100-000-medical-decisions",
    "videoId": "cZ5ZJy19KMo",
    "event": "AI Engineer Summit 2025",
    "durationMs": 735000,
    "speakers": [
      {
        "slug": "christopher-lovejoy",
        "name": "Christopher Lovejoy",
        "organization": "Anterior"
      }
    ],
    "topics": [
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Christopher Lovejoy explains how Anterior evaluates LLM-generated medical prior-authorization decisions when clinical nuance makes mistakes unacceptable. He demonstrates clinician review through the internal Scalpel dashboard, explains why fixed-rate manual review becomes impractical as decision volume grows, and describes combining expert-generated ground truths with reference-free LLM-as-judge evaluation and confidence scoring. Production integration routes uncertain cases to stronger models or clinicians, improving alignment with human reviewers while supporting timely customer responses."
  },
  {
    "slug": "model-maxxing-rft-dpo-sft-fine-tuning-with-openai-ilan-bigio-openai",
    "title": "Model-Maxxing: RFT, DPO, SFT (Fine-tuning with OpenAI) — Ilan Bigio, OpenAI",
    "url": "https://ai.engineer/talks/model-maxxing-rft-dpo-sft-fine-tuning-with-openai-ilan-bigio-openai",
    "videoId": "JfaLQqfXqPA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 6375000,
    "speakers": [
      {
        "slug": "ilan-bigio",
        "name": "Ilan Bigio",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "In an interactive OpenAI workshop, Ilan Bigio introduces supervised fine-tuning, direct preference optimization, and reinforcement fine-tuning, then explores synthetic training data, constrained classification and function-calling tasks, and model distillation using GPT-3.5 and GPT-4. Audience questions examine when fine-tuning is warranted, executable code evaluations, reinforcement fine-tuning for agents and reasoning models, o3 versus o1, and alternatives such as retrieval-augmented generation."
  },
  {
    "slug": "modern-post-training-a-deep-dive-will-brown-prime-intellect",
    "title": "Modern Post-Training: A Deep Dive — Will Brown, Prime Intellect",
    "url": "https://ai.engineer/talks/modern-post-training-a-deep-dive-will-brown-prime-intellect",
    "videoId": "V-EDrhIhHzQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2812000,
    "speakers": [
      {
        "slug": "will-brown",
        "name": "Will Brown",
        "organization": "Prime Intellect"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Prime Intellect applied-research lead Will Brown presents a workshop on open-source modern post-training infrastructure, centered on verifiers environments and the prime-rl asynchronous reinforcement-learning framework. He explains how shared environments and rollouts support evaluation, reinforcement learning, supervised fine-tuning, and expert distillation, and describes model-selection economics, the Environments Hub, and Lab's hosted training, inference, and sandbox capabilities."
  },
  {
    "slug": "monetizing-ai-alvaro-morales-orb",
    "title": "Monetizing AI — Alvaro Morales, Orb",
    "url": "https://ai.engineer/talks/monetizing-ai-alvaro-morales-orb",
    "videoId": "6WQYLQB0odc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1098000,
    "speakers": [
      {
        "slug": "alvaro-morales",
        "name": "Alvaro Morales",
        "organization": "Orb"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Orb co-founder and CEO Alvaro Morales explains why volatile inference costs, margin pressure, and customer demands for measurable ROI make AI products difficult to monetize. He compares direct and indirect monetization and traces AI-agent pricing from tokens and tasks to outcomes, using GitHub Copilot, Notion AI, Expedia, Vercel's v0, Zapier, and Intercom's Fin as examples. He also demonstrates selecting a time period and customer cohort for pricing simulations."
  },
  {
    "slug": "moondream-how-does-a-tiny-vision-model-slap-so-hard",
    "title": "Moondream: how does a tiny vision model slap so hard?",
    "url": "https://ai.engineer/talks/moondream-how-does-a-tiny-vision-model-slap-so-hard",
    "videoId": "T7sxvrJLJ14",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1166000,
    "speakers": [
      {
        "slug": "vikhyat-korrapati",
        "name": "Vikhyat Korrapati",
        "organization": "M87 Labs"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Vikhyat Korrapati of M87 Labs explains how Moondream, an open-source vision-language model with fewer than two billion parameters, performs image question answering, captioning, object localization, and counting while comparing favorably with larger models on VQA v2 and GQA. He describes its origins in browser-testing automation, its combination of pretrained SigLIP and Phi-1.5 components, and why poorly matched GPT-4-generated synthetic training data can increase hallucinations. The talk also addresses spatial reasoning, representative training distributions, enterprise deployment flexibility, and practical computer-vision applications."
  },
  {
    "slug": "most-enterprise-agentic-projects-are-doomed-here-s-why",
    "title": "Most Enterprise Agentic Projects Are Doomed — Here’s Why",
    "url": "https://ai.engineer/talks/most-enterprise-agentic-projects-are-doomed-here-s-why",
    "videoId": "AGkzpxMdPn8",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1235000,
    "speakers": [
      {
        "slug": "jess-grogan-avignon",
        "name": "Jess Grogan-Avignon",
        "organization": "Accenture"
      },
      {
        "slug": "jack-wang",
        "name": "Jack Wang",
        "organization": "Accenture"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "finance",
        "name": "Finance"
      }
    ],
    "summary": "Accenture presenters Jess Grogan-Avignon and Jack Wang argue that enterprise agentic-AI projects fail primarily because governance, funding, delivery, and deployment processes still operate at human speed. Drawing on Accenture’s AI-maturity research and a project delayed twelve months by cross-team approvals, they recommend investing in engineering automation and CI/CD, treating AI initiatives as a portfolio of bets, iterating on evidence, and building trust through human-supervised advisory modes and continuous feedback."
  },
  {
    "slug": "move-fast-break-nothing",
    "title": "Move Fast Break Nothing",
    "url": "https://ai.engineer/talks/move-fast-break-nothing",
    "videoId": "oJz4wveQ4oM",
    "event": "AI Engineer Summit 2023",
    "durationMs": 781000,
    "speakers": [
      {
        "slug": "dedy-kredo",
        "name": "Dedy Kredo",
        "organization": "CodiumAI"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Dedy Kredo presents CodiumAI's GAN-inspired vision of pairing AI code generation with a critic that checks developer intent, code behavior and edge cases. In a VS Code demonstration using the untested AutoScraper project, he generates and iteratively repairs tests, applies code-quality and performance suggestions including BLAKE3, and uses pull-request tooling for commit messages, review and security fixes. He concludes with personal remarks concerning violence and the Gaza Strip."
  },
  {
    "slug": "moving-away-from-agile-what-s-next",
    "title": "Moving away from Agile: What's Next?",
    "url": "https://ai.engineer/talks/moving-away-from-agile-what-s-next",
    "videoId": "SZStlIhyTCY",
    "event": "AI Engineer Code 2025",
    "durationMs": 1315000,
    "speakers": [
      {
        "slug": "martin-harrysson",
        "name": "Martin Harrysson",
        "organization": "McKinsey & Company"
      },
      {
        "slug": "natasha-maniar",
        "name": "Natasha Maniar",
        "organization": "McKinsey & Company"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "McKinsey’s Martin Harrysson and Natasha Maniar argue that AI-assisted software development requires redesigning operating models rather than simply adding coding tools to established Agile teams. Drawing on enterprise research, they describe end-to-end AI-native workflows, smaller multidisciplinary engineering pods, continuous planning, specification-driven collaboration with agents, and investments in upskilling, incentives, and organizational change. Cursor illustrates how agent-oriented workflows can reduce downstream rework and reshape engineering-team organization."
  },
  {
    "slug": "multi-agent-ai-and-network-knowledge-graphs-for-change-management-and-network-testing",
    "title": "Multi-Agent AI and Network Knowledge Graphs for Change Management and Network Testing",
    "url": "https://ai.engineer/talks/multi-agent-ai-and-network-knowledge-graphs-for-change-management-and-network-testing",
    "videoId": "m0dxZ-NDKHo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1129000,
    "speakers": [
      {
        "slug": "ola-mabadeje",
        "name": "Ola Mabadeje",
        "organization": "Cisco / Outshift by Cisco"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Cisco Outshift’s Ola Mabadeje presents an MVP for reducing failures in network change management by connecting ServiceNow ITSM tickets to cooperating AI agents and a live network knowledge graph. He explains heterogeneous YANG, JSON, and telemetry ingestion; graph performance and schema requirements; GraphRAG-style vector indexing; the choice of ArangoDB over Neo4j for the initial implementation; and automated test-plan generation for firewall changes. He concludes that knowledge graphs and an open agent-building framework are essential to scaling the approach."
  },
  {
    "slug": "multi-model-multimodal-and-multi-agent-innovations-in-azure-ai",
    "title": "Multi-model, multimodal, and multi-agent innovations in Azure AI",
    "url": "https://ai.engineer/talks/multi-model-multimodal-and-multi-agent-innovations-in-azure-ai",
    "videoId": "YcOrV-TLNyQ",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1736000,
    "speakers": [
      {
        "slug": "cedric-vidal",
        "name": "Cedric Vidal",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Microsoft principal AI advocate Cedric Vidal demonstrates Azure AI Studio capabilities spanning GPT-4o visual reasoning, Phi-3 and other model options, AI assistants, video translation, and serverless inference. Examples include extracting information from photographed menus, assessing infrastructure damage, analyzing GPS activity data, and generating a Java graphical frontend from a Python code repository."
  },
  {
    "slug": "multiplayer-agentic-engineering-enabling-your-whole-team-and-your-best-agents-to-work-together",
    "title": "Multiplayer agentic engineering: enabling your whole team and your best agents to work together",
    "url": "https://ai.engineer/talks/multiplayer-agentic-engineering-enabling-your-whole-team-and-your-best-agents-to-work-together",
    "videoId": "OL7kfezynJM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1124000,
    "speakers": [
      {
        "slug": "arjun-singh",
        "name": "Arjun Singh",
        "organization": "Superconductor"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Superconductor co-founder Arjun Singh explains how engineering teams can collaborate with coding agents through persistent shared sessions spanning Slack, GitHub, desktop, and mobile. He advocates model- and harness-agnostic workflows, automated conversion of meetings and customer feedback into pull requests, isolated cloud execution with controlled access, and project-specific agent evaluation and routing."
  },
  {
    "slug": "my-ai-thinks-i-m-eating-my-feelings-and-other-nutritional-insights",
    "title": "My AI Thinks I'm Eating My Feelings (and Other Nutritional Insights)",
    "url": "https://ai.engineer/talks/my-ai-thinks-i-m-eating-my-feelings-and-other-nutritional-insights",
    "videoId": "Ghc-qalQFLw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 599000,
    "speakers": [
      {
        "slug": "rami-alhamad",
        "name": "Rami Alhamad",
        "organization": "Alma"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Alma founder Rami Alhamad explains how his team builds a personalized AI nutrition companion, including an evidence-informed Alma Score developed with academic adviser Eric Rimm. He describes improving reliability by constraining large language models to narrow tasks, progressively returning food-recognition results to an iOS client while completing USDA database matching, and learning from beta feedback and tracking streaks. Additional product lessons include persistent user-editable context, proactive engagement, voice-based meal logging, and supporting multiple interaction modalities."
  },
  {
    "slug": "navigating-challenges-and-technical-debt-in-llms-deployment",
    "title": "Navigating Challenges and Technical Debt in LLMs Deployment",
    "url": "https://ai.engineer/talks/navigating-challenges-and-technical-debt-in-llms-deployment",
    "videoId": "IbJ40EwaNlM",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 975000,
    "speakers": [
      {
        "slug": "ahmed-menshawy",
        "name": "Ahmed Menshawy",
        "organization": "Mastercard"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Mastercard AI engineering leader Ahmed Menshawy describes the shift from structured-data machine learning to enterprise applications built around unstructured data and large language models. He argues for human augmentation and attention to present-day AI risks rather than speculative AGI scenarios, explains ChatGPT's conversational-interface breakthrough, and highlights Mastercard's generative-AI fraud-detection work. The talk then examines inference infrastructure, GPU and memory demands, system-level technical debt beyond model code, and the challenge of jointly optimizing retrieval and generation in RAG systems."
  },
  {
    "slug": "navigating-rag-optimization-with-an-evaluation-driven-compass",
    "title": "Navigating RAG Optimization with an Evaluation-Driven Compass",
    "url": "https://ai.engineer/talks/navigating-rag-optimization-with-an-evaluation-driven-compass",
    "videoId": "DId2KP8Ykz4",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1094000,
    "speakers": [
      {
        "slug": "atita-arora",
        "name": "Atita Arora",
        "organization": "Qdrant"
      },
      {
        "slug": "deanna-emery",
        "name": "Deanna Emery",
        "organization": "Quotient"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Atita Arora of Qdrant and Deanna Emery of Quotient explain how to improve retrieval-augmented generation through systematic evaluation rather than unnecessary architectural complexity. They compare naive and advanced RAG pipelines, discuss vector databases, query rewriting, re-ranking, context sizing, missing source information, scalability, and security, then show how domain-specific evaluation datasets and context-relevance metrics guide iterative experimentation and identify retrieval weaknesses."
  },
  {
    "slug": "neil-zeghidour-voice-ai-when-is-the-her-moment",
    "title": "Neil Zeghidour - Voice AI: when is the \"Her\" moment?",
    "url": "https://ai.engineer/talks/neil-zeghidour-voice-ai-when-is-the-her-moment",
    "videoId": "P_RI1kCkRbo",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1167000,
    "speakers": [
      {
        "slug": "neil-zeghidour",
        "name": "Neil Zeghidour",
        "organization": "Gradium AI"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "Gradium AI CEO Neil Zeghidour examines why contemporary voice agents still fall short of natural, Her-like conversation. Using voice-cloning and Reachy Mini demonstrations, he contrasts cascaded speech-to-text, LLM, and text-to-speech systems with speech-to-speech dialogue, emphasizing half-duplex limitations, simultaneous speech, and Moshi's full-duplex architecture. He also highlights paralinguistic understanding and introduces Gradium Phonon, an on-device text-to-speech model that runs locally on a CPU."
  },
  {
    "slug": "new-york-times-connections-a-case-study-on-nlp-in-word-games",
    "title": "New York Times' Connections: A Case Study on NLP in Word Games",
    "url": "https://ai.engineer/talks/new-york-times-connections-a-case-study-on-nlp-in-word-games",
    "videoId": "P_uhFGH4J9Y",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1111000,
    "speakers": [
      {
        "slug": "shafik-quoraishee",
        "name": "Shafik Quoraishee",
        "organization": "The New York Times Games"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "New York Times Games engineer Shafik Quoraishee presents independent, preliminary research into using Connections as a benchmark for AI reasoning. He explains how decoys, semantic ambiguity, and polysemy challenge language models, then describes modeling word relationships as a graph-coloring problem and applying a graph convolutional neural network to generate candidate solutions. He emphasizes that Connections puzzles are human-made and that his experiments are separate from internal New York Times research."
  },
  {
    "slug": "no-code-fine-tuning-mark-hennings",
    "title": "No-code Fine-tuning: Mark Hennings",
    "url": "https://ai.engineer/talks/no-code-fine-tuning-mark-hennings",
    "videoId": "LksXn4CLC0g",
    "event": "AI Engineer Summit 2023",
    "durationMs": 567000,
    "speakers": [
      {
        "slug": "mark-hennings",
        "name": "Mark Hennings",
        "organization": "Entry Point AI"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Entry Point AI cofounder Mark Hennings explains how no-code fine-tuning specializes large language models for practical business tasks and contrasts it with prompt engineering, discussing shorter prompts, performance, cost, larger training datasets, and collaboration. He demonstrates an Entry Point workflow that imports CSV examples, generates missing press-release inputs, combines GPT-3.5 Turbo fine-tuning with system prompts, and tests a trained model in Entry Point Playground."
  },
  {
    "slug": "no-more-bad-outputs-with-structured-generation",
    "title": "No more bad outputs with structured generation",
    "url": "https://ai.engineer/talks/no-more-bad-outputs-with-structured-generation",
    "videoId": "aNmfvN6S_n4",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 932000,
    "speakers": [
      {
        "slug": "remi-louf",
        "name": "Rémi Louf",
        "organization": ".txt (Outlines)"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Rémi Louf, co-founder of .txt and co-author of Outlines, explains how structured generation constrains language-model decoding to produce dependable outputs instead of relying on brittle prompting or post-processing. He demonstrates regex-guided answers, JSON Schema and Pydantic-controlled JSON, vision-model extraction, and integrations with open-model serving frameworks. He outlines the role of logits and next-token generation, then presents a Mistral 7B v0.1 GSM8K comparison suggesting structured one-shot prompting can approach the accuracy of unstructured eight-shot prompting."
  },
  {
    "slug": "no-more-slop-swyx",
    "title": "No More Slop – swyx",
    "url": "https://ai.engineer/talks/no-more-slop-swyx",
    "videoId": "IoiHI7p12Ao",
    "event": "AI Engineer Code 2025",
    "durationMs": 555000,
    "speakers": [
      {
        "slug": "swyx",
        "name": "Shawn \"swyx\" Wang",
        "organization": "AI Engineer"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "AI Engineer curator Shawn \"swyx\" Wang argues that slop means low-quality, inauthentic, or inaccurate work, whether produced by humans or AI. He connects the Code Summit's emphasis on curation to examples from AI-generated video, entertainment, startups, and vibe coding, then describes using AI itself to improve information quality, codebase understanding, semi-asynchronous development, computer-use automation, and modular human-AI collaboration."
  },
  {
    "slug": "notion-s-token-town",
    "title": "Notion's Token Town",
    "url": "https://ai.engineer/talks/notion-s-token-town",
    "videoId": "-I5W5QVAT8E",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1435000,
    "speakers": [
      {
        "slug": "sarah-sachs",
        "name": "Sarah Sachs",
        "organization": "Notion"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Notion AI engineering leader Sarah Sachs argues that sustainable AI products cannot depend on favorable per-token pricing: model upgrades can increase actual output costs, and relying on one provider creates strategic lock-in. She describes Notion as a durable system of record for collaboration among people and agents, traces the progression from assistants to interconnected AI systems, and advocates model-provider optionality, matching capabilities to task complexity, and using open-weight models where appropriate."
  },
  {
    "slug": "on-ai-and-knowledge",
    "title": "On AI and Knowledge",
    "url": "https://ai.engineer/talks/on-ai-and-knowledge",
    "videoId": "RGSFUqzqErE",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1055000,
    "speakers": [
      {
        "slug": "pablo-castro",
        "name": "Pablo Castro",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Microsoft corporate vice president and distinguished engineer Pablo Castro examines intrinsic model knowledge, extrinsic enterprise knowledge, and learned knowledge in AI agents. He traces the progression from IntelliSense and GitHub Copilot to modern coding models, outlines Microsoft Foundry and the Microsoft IQ family, and explains how Foundry IQ uses Azure AI Search and agentic retrieval to ground agents in organizational and public information. He closes by describing learning loops based on real agent execution traces."
  },
  {
    "slug": "on-curiosity-sharif-shameem-lexica",
    "title": "On Curiosity — Sharif Shameem, Lexica",
    "url": "https://ai.engineer/talks/on-curiosity-sharif-shameem-lexica",
    "videoId": "0F8mnGPUycY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1115000,
    "speakers": [
      {
        "slug": "sharif-shameem",
        "name": "Sharif Shameem",
        "organization": "Lexica"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "Lexica founder Sharif Shameem argues that curiosity and exploratory demos reveal AI capabilities before their practical applications are obvious. He revisits early GPT-3 experiments in browser-based interface generation and constrained-context web agents, highlights Gemini 2.5 Pro video analysis, and frames AI engineering as open-ended discovery rather than purely goal-directed implementation. An audience discussion follows the presentation."
  },
  {
    "slug": "on-engineering-ai-systems-that-endure-the-bitter-lesson",
    "title": "On Engineering AI Systems that Endure The Bitter Lesson",
    "url": "https://ai.engineer/talks/on-engineering-ai-systems-that-endure-the-bitter-lesson",
    "videoId": "qdmxApz3EJI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1152000,
    "speakers": [
      {
        "slug": "omar-khattab",
        "name": "Omar Khattab",
        "organization": "Databricks"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Omar Khattab examines how AI engineers can build durable software despite rapidly changing language models, APIs, and optimization techniques. Interpreting Rich Sutton's Bitter Lesson, he distinguishes scaling search and learning to improve intelligence from engineering reliable, purpose-built systems. He advocates avoiding premature, prompt-specific complexity and instead using separation of concerns, modular abstractions, interchangeable models, and optimizers that can apply reinforcement learning or prompt optimization across applications."
  },
  {
    "slug": "one-developer-two-dozen-agents-zero-alignment-why-we-need-collaborative-ai-engineering",
    "title": "One Developer, Two Dozen Agents, Zero Alignment: Why we Need Collaborative AI Engineering",
    "url": "https://ai.engineer/talks/one-developer-two-dozen-agents-zero-alignment-why-we-need-collaborative-ai-engineering",
    "videoId": "ClWD8OEYgp8",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1063000,
    "speakers": [
      {
        "slug": "maggie-appleton",
        "name": "Maggie Appleton",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "GitHub Next researcher Maggie Appleton argues that running many coding agents per developer increases output without solving team alignment, shared context, duplicated work, merge conflicts, or overloaded pull-request review. She demonstrates Ace, the Agent Collaboration Environment: an experimental multiplayer workspace combining shared conversations, sandboxed cloud microVMs, isolated Git branches, Claude Opus 4.6, collaborative VS Code editing, and GitHub pull-request integration so developers, designers, product managers, and other teammates can coordinate agent-assisted software development in real time."
  },
  {
    "slug": "one-login-to-rule-them-all-cross-app-access-for-mcp",
    "title": "One Login to Rule Them All: Cross-App Access for MCP",
    "url": "https://ai.engineer/talks/one-login-to-rule-them-all-cross-app-access-for-mcp",
    "videoId": "EmhRyw6xeT0",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1404000,
    "speakers": [
      {
        "slug": "garrett-galow",
        "name": "Garrett Galow",
        "organization": "WorkOS"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "WorkOS product leader Garrett Galow explains why MCP’s OAuth-based, server-by-server consent disrupts enterprise single sign-on and obscures IT oversight. He presents Cross-App Access, an identity-provider-governed approach using identity assertions and scoped token exchange, demonstrates Claude Code authenticating through Okta, and discusses authorization scopes, agent governance, and OIDC support during audience questions."
  },
  {
    "slug": "one-model-to-rule-recommendations-netflix-s-big-bet",
    "title": "One model to rule recommendations: Netflix's Big Bet",
    "url": "https://ai.engineer/talks/one-model-to-rule-recommendations-netflix-s-big-bet",
    "videoId": "AbZ4IYGbfpQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1348000,
    "speakers": [
      {
        "slug": "yesu-feng",
        "name": "Yesu Feng",
        "organization": "Netflix"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Netflix researcher Yesu Feng explains the company's strategy of replacing fragmented personalization models with a shared recommendation foundation model. He describes diverse homepage and content-recommendation requirements, autoregressive transformer training on multifield user-interaction events, tokenization choices, combining ID and semantic embeddings to handle cold starts, multi-token prediction, shared representation and embedding infrastructure, and generative retrieval for collections. The recording concludes with moderated audience questions."
  },
  {
    "slug": "one-registry-to-rule-them-all-sonny-merla-mauro-luchetti-mattia-redaelli-quantyca",
    "title": "One Registry to Rule them All - Sonny Merla, Mauro Luchetti, & Mattia Redaelli, Quantyca",
    "url": "https://ai.engineer/talks/one-registry-to-rule-them-all-sonny-merla-mauro-luchetti-mattia-redaelli-quantyca",
    "videoId": "VXfRt_H-V08",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1367000,
    "speakers": [
      {
        "slug": "sonny-merla",
        "name": "Sonny Merla",
        "organization": "Amplifon"
      },
      {
        "slug": "mauro-luchetti",
        "name": "Mauro Luchetti",
        "organization": "Quantyca"
      },
      {
        "slug": "mattia-redaelli",
        "name": "Mattia Redaelli",
        "organization": "Quantyca"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Sonny Merla of Amplifon and Mauro Luchetti and Mattia Redaelli of Quantyca explain how Amplifon's AmplifAI program addresses enterprise AI governance and fragmented agent development. Their architecture combines a centralized AI gateway for model access, Microsoft Entra ID authentication, budgeting and auditing with registries for MCP servers and A2A agents. Agent cards, use-case linkage and object lineage make deployed capabilities discoverable, traceable and easier to govern across teams."
  },
  {
    "slug": "open-challenges-for-ai-engineering",
    "title": "Open Challenges for AI Engineering",
    "url": "https://ai.engineer/talks/open-challenges-for-ai-engineering",
    "videoId": "eTTMUWP5B0s",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1129000,
    "speakers": [
      {
        "slug": "simon-willison",
        "name": "Simon Willison"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Simon Willison argues that GPT-4's period of dominance has ended as competing frontier, inexpensive, and open models expand engineers' choices. He contrasts MMLU with human-preference evaluations in Chatbot Arena, demonstrates AI-assisted visualizations, warns that prompt injection can exfiltrate private data through Markdown images across chatbot products, and calls for accountability over unrequested, unreviewed AI-generated content."
  },
  {
    "slug": "open-questions-for-ai-engineering",
    "title": "Open Questions for AI Engineering",
    "url": "https://ai.engineer/talks/open-questions-for-ai-engineering",
    "videoId": "AjLVoAu-u-Q",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1473000,
    "speakers": [
      {
        "slug": "simon-willison",
        "name": "Simon Willison"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Simon Willison reviews the emergence of AI engineering, tracing developments from GPT-3 and ChatGPT to GPT-4, Bing Chat, and commercially usable local models such as Llama 2. He demonstrates his LLM command-line tooling and a Claude-powered Hacker News summarizer backed by SQLite, while examining conversational interfaces, prompt-injection risks around private data, AI-assisted programming, and WebAssembly as a possible sandboxing approach."
  },
  {
    "slug": "open-rag-eval-rag-evaluation-without-golden-answers",
    "title": "open-rag-eval: RAG Evaluation without \"golden\" answers.",
    "url": "https://ai.engineer/talks/open-rag-eval-rag-evaluation-without-golden-answers",
    "videoId": "1cQlnfwmIdU",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 303000,
    "speakers": [
      {
        "slug": "ofer-mendelevitch",
        "name": "Ofer Mendelevitch",
        "organization": "Vectara"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Vectara's Ofer Mendelevitch introduces open-rag-eval, an open-source framework for evaluating RAG pipelines without golden answers or reference chunks. He explains its connector-based architecture and four evaluation areas: UMBRELA retrieval relevance scoring, AutoNuggetizer answer assessment with LLM-judged information nuggets, citation faithfulness, and HHEM-based hallucination detection. The presentation also demonstrates evaluation-result visualization and integrations with Vectara, LangChain, and LlamaIndex."
  },
  {
    "slug": "open-source-is-dead-long-live-open-source",
    "title": "Open Source Is Dead. Long Live Open Source.",
    "url": "https://ai.engineer/talks/open-source-is-dead-long-live-open-source",
    "videoId": "CoEIs6Xm8m8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1050000,
    "speakers": [
      {
        "slug": "saoud-rizwan",
        "name": "Saoud Rizwan",
        "organization": "Cline"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Cline founder Saoud Rizwan argues that AI-generated contributions, maintainer overload, and software supply-chain compromises are undermining traditional open-source collaboration. Using examples involving Zig, curl, tldraw, and LiteLLM, he contrasts those risks with escalating proprietary coding-agent costs and vendor lock-in, then argues that auditable agents, open-weight models, and competition among inference providers can restore developer choice and reduce costs."
  },
  {
    "slug": "openai-for-vps-of-ai",
    "title": "OpenAI for VPs of AI",
    "url": "https://ai.engineer/talks/openai-for-vps-of-ai",
    "videoId": "joHR2pmxDQE",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1012000,
    "speakers": [
      {
        "slug": "prashant-mital",
        "name": "Prashant Mital",
        "organization": "OpenAI"
      },
      {
        "slug": "toki-sherbakov",
        "name": "Toki Sherbakov",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "OpenAI presenters Prashant Mital and Toki Sherbakov describe an enterprise AI adoption progression from workforce enablement to internal automation and customer-facing products, emphasizing business strategy, measurable outcomes, organizational enablement, and hands-on implementation support. They then outline practical agent-building guidance: limit abstractions, begin with one agent, expand to agent networks only when complexity demands it, keep prompts focused, and use guardrails for edge cases."
  },
  {
    "slug": "openclaw-in-your-hand-building-a-physical-ai-terminal-for-local-llm-agents",
    "title": "OpenClaw in Your Hand: Building a Physical AI Terminal for Local LLM Agents",
    "url": "https://ai.engineer/talks/openclaw-in-your-hand-building-a-physical-ai-terminal-for-local-llm-agents",
    "videoId": "akk6KRlcwW4",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1476000,
    "speakers": [
      {
        "slug": "lech-kalinowski",
        "name": "Lech Kalinowski",
        "organization": "Callstack"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Callstack engineer Lech Kalinowski presents Vault, a physical, text-first handheld terminal that controls OpenClaw agents through an ESP32-based device and separate local backend. He explains its dual-display interface, constrained firmware and pre-allocated memory, local GPT-model inference on NVIDIA DGX Spark with TensorRT-LLM, and an experimental narrative-focused RPG mode."
  },
  {
    "slug": "openllmetry-is-all-you-need",
    "title": "OpenLLMetry is all you need",
    "url": "https://ai.engineer/talks/openllmetry-is-all-you-need",
    "videoId": "KVgbERRPU4M",
    "event": "AI Engineer Summit 2025",
    "durationMs": 552000,
    "speakers": [
      {
        "slug": "nir-gazit",
        "name": "Nir Gazit",
        "organization": "Traceloop"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Traceloop co-founder and CEO Nir Gazit introduces OpenLLMetry, an open-source extension of OpenTelemetry for observing generative-AI applications. He explains OpenTelemetry logs, metrics, traces, SDKs, automatic instrumentation, and collectors, including filtering sensitive data. He then shows how OpenLLMetry instruments model providers, vector databases, and frameworks, illustrates Pinecone query and vector tracing, and emphasizes portability across existing observability platforms."
  },
  {
    "slug": "openrag-an-open-source-stack-for-rag-phil-nash",
    "title": "OpenRAG: An open-source stack for RAG — Phil Nash",
    "url": "https://ai.engineer/talks/openrag-an-open-source-stack-for-rag-phil-nash",
    "videoId": "4TxOBhDRRCM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 952000,
    "speakers": [
      {
        "slug": "phil-nash",
        "name": "Phil Nash",
        "organization": "IBM"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "IBM developer relations engineer Phil Nash introduces OpenRAG, an open-source RAG stack combining Docling document processing, OpenSearch retrieval, and Langflow orchestration. He explains why large context windows do not eliminate retrieval challenges, demonstrates document-ingestion pipelines including OCR, automatic speech recognition, hierarchical chunking, and the Granite-Docling-258M vision-language model, and discusses configurable embeddings, offline deployment, local demonstrations, and application customization."
  },
  {
    "slug": "openthoughts-data-recipes-for-reasoning-models",
    "title": "OpenThoughts: Data Recipes for Reasoning Models",
    "url": "https://ai.engineer/talks/openthoughts-data-recipes-for-reasoning-models",
    "videoId": "liG97YXaTSA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1199000,
    "speakers": [
      {
        "slug": "ryan-marten",
        "name": "Ryan Marten",
        "organization": "Bespoke Labs"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Bespoke Labs founding engineer Ryan Marten presents OpenThoughts, an open effort to develop reasoning datasets and models through supervised fine-tuning. He introduces OpenThoughts3, discusses evaluation on AIME, LiveCodeBench, and GPQA Diamond, and explains how dataset recipes, scaling, domain-specific filtering, and teacher selection affect performance across mathematics, coding, and science. The recording concludes with moderator-led audience questions about supervised fine-tuning, teacher models, and reasoning-trace failures."
  },
  {
    "slug": "optimizing-inference-for-voice-models-in-production",
    "title": "Optimizing inference for voice models in production",
    "url": "https://ai.engineer/talks/optimizing-inference-for-voice-models-in-production",
    "videoId": "gmTHs5T_YAE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 913000,
    "speakers": [
      {
        "slug": "philip-kiely",
        "name": "Philip Kiely",
        "organization": "Baseten"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Baseten's Philip Kiely explains how to optimize production text-to-speech inference using Canopy Labs' open-source Orpheus TTS, which builds on a Llama 3B backbone. He contrasts maximizing LLM token throughput with meeting Orpheus's approximately 83-token-per-second real-time threshold, then prioritizing time to first byte, concurrent streams, and GPU efficiency. Techniques include TensorRT-LLM, FP8 quantization, GPU-based audio decoding with torch.compile and PyTorch inference mode, and attention to infrastructure and client-side latency."
  },
  {
    "slug": "optimizing-llms-in-insurance-with-dspy-beyond-manual-tuning",
    "title": "Optimizing LLMs in Insurance with DSPy: Beyond Manual Tuning",
    "url": "https://ai.engineer/talks/optimizing-llms-in-insurance-with-dspy-beyond-manual-tuning",
    "videoId": "IAdZxqjZ45U",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1169000,
    "speakers": [
      {
        "slug": "jeronim-morina",
        "name": "Jeronim Morina",
        "organization": "AXA"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "AXA engineer Jeronim Morina argues that production insurance applications need first-principles engineering, meaningful evaluation, and modular LLM programs rather than brittle manual prompt tuning. He describes AXA Germany's data organization and Secure GPT platform, discusses OpenAI, Azure, and Mistral AI, compares structured-output and prompting tools, recommends inspecting generated prompts with mitmproxy, and presents DSPy as a framework for optimizing modular language-model systems."
  },
  {
    "slug": "paperclip-open-source-human-control-plane-for-ai-labor-dotta",
    "title": "Paperclip: Open Source Human Control Plane for AI Labor — Dotta",
    "url": "https://ai.engineer/talks/paperclip-open-source-human-control-plane-for-ai-labor-dotta",
    "videoId": "h403btjldDQ",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1474000,
    "speakers": [
      {
        "slug": "dotta",
        "name": "Dotta",
        "organization": "Paperclip"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Paperclip creator Dotta demonstrates an open-source control plane for organizing and supervising AI-agent work. He shows installation with npx paperclipai onboard, hierarchical delegation through CEO, CTO, and specialist agents, interoperability with Codex, Claude, Gemini, and OpenClaw, and agent-managed hiring, skills, shared context, and Remotion video production. Further examples cover recurring research routines, Greptile-assisted code review, human approval of hiring plans, spending visibility, and planned workflow improvements."
  },
  {
    "slug": "paying-engineers-like-salespeople",
    "title": "Paying Engineers like Salespeople",
    "url": "https://ai.engineer/talks/paying-engineers-like-salespeople",
    "videoId": "4mRekpZpBZs",
    "event": "AI Engineer Code 2025",
    "durationMs": 893000,
    "speakers": [
      {
        "slug": "arman-hezarkhani",
        "name": "Arman Hezarkhani",
        "organization": "Tenex"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Tenex co-founder Arman Hezarkhani explains an outcome-based compensation system that charges clients for delivered story points and pays engineers for completed story points. He argues that traditional compensation discourages AI adoption, describes the complementary strategist and AI-engineer roles used to deliver client roadmaps, and illustrates AI applications including retail traffic heat mapping while emphasizing hiring quality and aligned incentives."
  },
  {
    "slug": "perception-agents-antje-barth-amazon-agi-lab",
    "title": "Perception Agents — Antje Barth, Amazon AGI Lab",
    "url": "https://ai.engineer/talks/perception-agents-antje-barth-amazon-agi-lab",
    "videoId": "2JX6JYyQG4Y",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1305000,
    "speakers": [
      {
        "slug": "antje-barth",
        "name": "Antje Barth",
        "organization": "Amazon AGI Lab"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Amazon AGI Lab's Antje Barth argues that reliable agents need more than browser clicks, form filling, and API calls: they must perceive rendered interfaces, share visual context with people, and verify their own work. She introduces an open-source perception agent harness featuring browser-based annotation and visual verification, discusses interactions beyond the screen, and invites developers to experiment with the team's GitHub repositories."
  },
  {
    "slug": "perceptual-evaluations-evals-for-aesthetics-diego-rodriguez-krea-ai",
    "title": "Perceptual Evaluations: Evals for Aesthetics — Diego Rodriguez, Krea.ai",
    "url": "https://ai.engineer/talks/perceptual-evaluations-evals-for-aesthetics-diego-rodriguez-krea-ai",
    "videoId": "h5ItAJuB3Fc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 988000,
    "speakers": [
      {
        "slug": "diego-rodriguez",
        "name": "Diego Rodriguez",
        "organization": "Krea"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Krea cofounder Diego Rodriguez examines why generative-media evaluations often fail to capture human perception and aesthetic judgment. Drawing on information theory and perceptual compression in JPEG, MP3, and MP4, he shows how human sensory limits shape what is preserved or discarded. He then critiques FID and CLIP-style proxy metrics, highlighting clean-fid research showing that visually subtle JPEG artifacts can distort benchmark scores, and argues for evaluations that better learn subjective human preferences."
  },
  {
    "slug": "persona-engineering-a-field-guide-to-ai-synthetic-personas-ishan-anand-insightsciences-ai",
    "title": "Persona Engineering: A Field Guide to AI Synthetic Personas — Ishan Anand, InsightSciences.ai",
    "url": "https://ai.engineer/talks/persona-engineering-a-field-guide-to-ai-synthetic-personas-ishan-anand-insightsciences-ai",
    "videoId": "YnNF55QV0zs",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1269000,
    "speakers": [
      {
        "slug": "ishan-anand",
        "name": "Ishan Anand",
        "organization": "Insight Sciences"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Insight Sciences Chief AI Officer Ishan Anand examines how large language models can simulate research respondents while warning that synthetic personas behave more like forecasts than independent human observations. He discusses historical simulation attempts, interview-grounded agents, missing-context confounders, prompt sensitivity, the gap between stated attitudes and behavior, and bias introduced by persona construction. He highlights SubPOP-style fine-tuning against real survey-response distributions and explains why repeatedly sampling an unchanged forecast does not create additional certainty."
  },
  {
    "slug": "personality-driven-development-exploring-the-frontier-of-agents-with-attitude",
    "title": "Personality-Driven Development: Exploring the Frontier of Agents with Attitude",
    "url": "https://ai.engineer/talks/personality-driven-development-exploring-the-frontier-of-agents-with-attitude",
    "videoId": "lG_8rgQqxfc",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1080000,
    "speakers": [
      {
        "slug": "benjamin-stein",
        "name": "Benjamin Stein",
        "organization": "Perpetual"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Perpetual's Benjamin Stein explains how giving AI agents recognizable personalities, voices, forms, and workplace roles makes their capabilities easier to understand while raising customer expectations and introducing gender-related design biases. He argues that specialized virtual teammates also offer a practical engineering decomposition: narrower inputs and responsibilities reduce opportunities for LLM and tool-calling errors. The talk concludes with a vision of customizable AI employees whose behavior can be molded to individual business needs instead of fixed SaaS product assumptions."
  },
  {
    "slug": "personalization-in-the-era-of-llms",
    "title": "Personalization in the Era of LLMs",
    "url": "https://ai.engineer/talks/personalization-in-the-era-of-llms",
    "videoId": "5YSJEP0HWzM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1212000,
    "speakers": [
      {
        "slug": "shivam-verma",
        "name": "Shivam Verma",
        "organization": "Spotify"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Spotify AI Foundation technical lead Shivam Verma explains how Spotify combines user representations, catalog embeddings, hierarchical semantic content tokens, and open-weight language models to build steerable music and podcast recommendation systems. He connects this architecture to personalized products including Discover Weekly, DJ, and Prompted Playlist, and describes using listening history and tokenized content to generate the next relevant track or podcast episode."
  },
  {
    "slug": "piloting-agents-in-github-copilot",
    "title": "Piloting agents in GitHub Copilot",
    "url": "https://ai.engineer/talks/piloting-agents-in-github-copilot",
    "videoId": "DdaAABdAqZY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 3547000,
    "speakers": [
      {
        "slug": "christopher-harrison",
        "name": "Christopher Harrison",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Christopher Harrison leads a practical GitHub Copilot workshop covering GitHub organization access, coding agents that work from issues, custom instructions and coding standards, and Model Context Protocol connections to external tools. Audience questions address organization-level governance, agent firewall restrictions, GitHub Actions, testing, and preserving established DevOps practices."
  },
  {
    "slug": "pipecat-cloud-enterprise-voice-agents-built-on-open-source",
    "title": "Pipecat Cloud: Enterprise Voice Agents Built On Open Source",
    "url": "https://ai.engineer/talks/pipecat-cloud-enterprise-voice-agents-built-on-open-source",
    "videoId": "IA4lZjh9sTs",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1606000,
    "speakers": [
      {
        "slug": "kwindla-hultman-kramer",
        "name": "Kwindla Hultman Kramer",
        "organization": "Daily"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Daily co-founder Kwindla Hultman Kramer explains how to build responsive enterprise voice agents with the open-source, vendor-neutral Pipecat framework and deploy them on Pipecat Cloud. He highlights an 800-millisecond voice-to-voice latency target, turn detection, interruptions, asynchronous tools, telephony integrations, noise cancellation, evaluation, and observability. Audience questions extend the discussion to Kyutai's full-duplex Moshi model, Gemini Flash, Sesame, and the challenges of training audio-native conversational systems."
  },
  {
    "slug": "platforms-for-humans-and-machines-engineering-for-the-age-of-agents-juan-herreros-elorza",
    "title": "Platforms for Humans and Machines: Engineering for the Age of Agents — Juan Herreros Elorza",
    "url": "https://ai.engineer/talks/platforms-for-humans-and-machines-engineering-for-the-age-of-agents-juan-herreros-elorza",
    "videoId": "cCRO3ChaYhM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1275000,
    "speakers": [
      {
        "slug": "juan-herreros-elorza",
        "name": "Juan Herreros Elorza",
        "organization": "Banking Circle"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Banking Circle team lead Juan Herreros Elorza explains how its Atlas platform illustrates a broader principle: internal developer platforms should serve AI agents as effectively as human developers. He recommends API-based self-service, CLI or MCP interfaces, local validation and iterative execution, machine-accessible observability, discoverable documentation, and carefully governed AI-assisted contributions."
  },
  {
    "slug": "playground-in-prod-optimising-agents-in-production-environments",
    "title": "Playground in Prod - Optimising Agents in Production Environments",
    "url": "https://ai.engineer/talks/playground-in-prod-optimising-agents-in-production-environments",
    "videoId": "A48uhxfxbsM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 4840000,
    "speakers": [
      {
        "slug": "samuel-colvin",
        "name": "Samuel Colvin",
        "organization": "Pydantic"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Pydantic founder Samuel Colvin demonstrates how to improve deployed AI agents using Pydantic AI, Logfire, evaluation datasets, GEPA-driven prompt optimization, and managed variables. A political-relationship extraction example illustrates golden-answer evaluation, Pareto-frontier candidate selection, agent tooling, observability, and practical tradeoffs involving model latency and prompt size. Audience questions accompany the single scheduled-speaker workshop."
  },
  {
    "slug": "poc-to-prod-hard-lessons-from-200-enterprise-genai-deployments",
    "title": "POC to PROD: Hard Lessons from 200+ Enterprise GenAI Deployments",
    "url": "https://ai.engineer/talks/poc-to-prod-hard-lessons-from-200-enterprise-genai-deployments",
    "videoId": "vW8wLsb3Nnc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1156000,
    "speakers": [
      {
        "slug": "randall-hunt",
        "name": "Randall Hunt",
        "organization": "Caylent"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Randall Hunt of Caylent shares practical lessons from enterprise generative AI deployments, including building-decarbonization agents, searchable wildlife-video archives, sports-highlight extraction, enterprise tool governance, and intelligent document processing. He describes multimodal indexing with Amazon Nova Pro, Titan embeddings, Elasticsearch, Postgres, and pgvector; improving sports-video analysis through audio cues and SAM 2-assisted visual annotations; and balancing AWS infrastructure choices, prompt maintenance, and production economics."
  },
  {
    "slug": "practical-graphrag-making-llms-smarter-with-knowledge-graphs",
    "title": "Practical GraphRAG: Making LLMs Smarter with Knowledge Graphs",
    "url": "https://ai.engineer/talks/practical-graphrag-making-llms-smarter-with-knowledge-graphs",
    "videoId": "XNneh6-eyPg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1186000,
    "speakers": [
      {
        "slug": "michael-hunger",
        "name": "Michael Hunger",
        "organization": "Neo4j"
      },
      {
        "slug": "stephen-chin",
        "name": "Stephen Chin",
        "organization": "Neo4j"
      },
      {
        "slug": "jesus-barrasa",
        "name": "Jesús Barrasa",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Neo4j presenters Michael Hunger and Stephen Chin explain how GraphRAG augments conventional retrieval-augmented generation with knowledge graphs, connected enterprise context, and graph-based retrieval to improve relevance, explainability, and factual grounding. They cover the limitations of vector-only retrieval, nodes and relationships, lexical graph modeling, schema-guided extraction, a Stack Overflow graph example, and Microsoft Research's GraphRAG work. Jesús Barrasa appears in the published session roster but is not independently identified in the supplied transcript excerpts."
  },
  {
    "slug": "practical-tactics-to-build-reliable-ai-apps-dmitry-kuchin-multinear",
    "title": "Practical tactics to build reliable AI apps — Dmitry Kuchin, Multinear",
    "url": "https://ai.engineer/talks/practical-tactics-to-build-reliable-ai-apps-dmitry-kuchin-multinear",
    "videoId": "-T6uZYYzkWw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 895000,
    "speakers": [
      {
        "slug": "dmitry-kuchin",
        "name": "Dmitry Kuchin",
        "organization": "Multinear"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Dmitry Kuchin describes how to turn unreliable generative-AI proofs of concept into production-ready applications by reverse-engineering evaluations from real user scenarios and business outcomes. He demonstrates using o3 to generate scenario-specific questions, checklists, and persona variations; introduces the open-source Multinear evaluation platform; and explains how iterative tests support model and architecture comparisons, LLM-as-a-judge workflows, and mock-database testing for Text-to-SQL."
  },
  {
    "slug": "pragmatic-ai-with-typechat",
    "title": "Pragmatic AI With TypeChat",
    "url": "https://ai.engineer/talks/pragmatic-ai-with-typechat",
    "videoId": "UNKDAo8YNmg",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1114000,
    "speakers": [
      {
        "slug": "daniel-rosenwasser",
        "name": "Daniel Rosenwasser",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Daniel Rosenwasser presents Microsoft’s TypeChat as a pragmatic bridge between natural-language model output and the structured data expected by conventional applications. He demonstrates using TypeScript interfaces to guide JSON generation, compiler diagnostics to validate and repair responses, and repository examples including sentiment classification and a coffee-shop JSON translator. The talk also explores representing multi-step operations as constrained JSON programs, reducing problems associated with unrestricted generated code, sandboxing, and API orchestration."
  },
  {
    "slug": "privacy-first-enterprise-ai-building-ai-agents-that-never-leave-your-security-boundary",
    "title": "Privacy First Enterprise AI: Building AI Agents that Never Leave Your Security Boundary",
    "url": "https://ai.engineer/talks/privacy-first-enterprise-ai-building-ai-agents-that-never-leave-your-security-boundary",
    "videoId": "_pBfv1rbLBU",
    "event": "AI Engineer Summit 2025",
    "durationMs": 430000,
    "speakers": [
      {
        "slug": "steven-moon",
        "name": "Steven Moon",
        "organization": "Aech AI"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Aech AI founder Steven Moon argues that enterprise AI agents should operate inside existing organizational security boundaries rather than through standalone SaaS portals. He describes using private-cloud infrastructure, Microsoft 365, Azure, Active Directory, and established security policies to provision, monitor, and audit agents like employees. Email enables observable agent-to-agent collaboration, while comparable patterns can also extend to Google Workspace and other existing enterprise platforms."
  },
  {
    "slug": "privacy-preserving-intelligence-steve-korshakov-bee-acq-amazon",
    "title": "Privacy-Preserving Intelligence — Steve Korshakov, Bee (acq. Amazon)",
    "url": "https://ai.engineer/talks/privacy-preserving-intelligence-steve-korshakov-bee-acq-amazon",
    "videoId": "IvE8n-ylFYY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 953000,
    "speakers": [
      {
        "slug": "steve-korshakov",
        "name": "Steve Korshakov",
        "organization": "Bee, Amazon"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Bee founding engineer Steve Korshakov explains how an Amazon-owned wearable personal AI protects highly sensitive conversational data through customer-device-managed encryption keys, independently verifiable Sigstore transparency logs, workload attestation, in-house inference, and certificates embedding attestation proofs. He describes private certificate authorities and EC2-based infrastructure, followed by audience questions about deployment security, auditing, and potentially unsafe AI agents."
  },
  {
    "slug": "proactive-agents",
    "title": "Proactive Agents",
    "url": "https://ai.engineer/talks/proactive-agents",
    "videoId": "v3u8xc0zLec",
    "event": "AI Engineer Code 2025",
    "durationMs": 1011000,
    "speakers": [
      {
        "slug": "kath-korevec",
        "name": "Kath Korevec",
        "organization": "Google Labs"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Google Labs product leader Kath Korevec argues that asynchronous coding agents still impose a mental burden when developers must initiate, monitor, and coordinate every task. She describes Jules evolving into a proactive teammate that understands repositories, remembers preferences, proposes maintenance work with visible rationale, and connects coding, design, and behavioral signals through tools such as Stitch and Insights. A walkthrough of codebase indexing and an embedded-hardware example illustrate how proactive agents could preserve developer attention for creative work."
  },
  {
    "slug": "production-evals-for-agentic-ai-systems",
    "title": "Production Evals For Agentic AI Systems",
    "url": "https://ai.engineer/talks/production-evals-for-agentic-ai-systems",
    "videoId": "vljxQZfJ9wY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 492000,
    "speakers": [
      {
        "slug": "nishant-gupta",
        "name": "Nishant Gupta",
        "organization": "Meta"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Meta software engineering tech lead Nishant Gupta explains why offline benchmarks and answer accuracy fail to capture agentic systems’ production behavior. He advocates evaluating complete workflows, tool use, failure recovery, and scenario-driven task completion using SRE-style reliability metrics, distributed tracing, production telemetry, continuous evaluation, and operational control planes."
  },
  {
    "slug": "production-software-keeps-breaking-and-it-will-only-get-worse-here-s-how-traversal-is-fixing-it",
    "title": "Production software keeps breaking, and it will only get worse. Here's how Traversal is fixing it.",
    "url": "https://ai.engineer/talks/production-software-keeps-breaking-and-it-will-only-get-worse-here-s-how-traversal-is-fixing-it",
    "videoId": "L6_NiGIEXZQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1093000,
    "speakers": [
      {
        "slug": "anish-agarwal",
        "name": "Anish Agarwal",
        "organization": "Traversal"
      },
      {
        "slug": "matthew-schoenbauer",
        "name": "Matthew Schoenbauer",
        "organization": "Traversal"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Traversal presenters Anish Agarwal and Matthew Schoenbauer argue that AI-generated software will increase production complexity while reducing engineers' operational context. They explain why observability dashboards, conventional AIOps, standalone LLMs, and runbook-dependent agents fall short, then describe an autonomous troubleshooting approach combining causal machine learning, semantic reasoning, and coordinated agent swarms. A DigitalOcean case study illustrates incident investigation and reduced mean time to resolution."
  },
  {
    "slug": "productionizing-genai-models-lessons-from-the-world-s-best-ai-teams",
    "title": "Productionizing GenAI Models – Lessons from the world's best AI teams",
    "url": "https://ai.engineer/talks/productionizing-genai-models-lessons-from-the-world-s-best-ai-teams",
    "videoId": "uiq95JYpBGY",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1356000,
    "speakers": [
      {
        "slug": "lukas-biewald",
        "name": "Lukas Biewald",
        "organization": "Weights & Biases"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Weights & Biases cofounder Lukas Biewald explains why impressive generative-AI demos are substantially easier to build than dependable production applications. Drawing on customer experience and an Alexa-style project fine-tuned with QLoRA on consumer hardware, he argues for combining prompt engineering, fine-tuning, evaluation, passive experiment tracking, reproducibility, and application-specific metrics. He closes by introducing W&B Weave as tooling for bringing AI applications into production."
  },
  {
    "slug": "prompt-engineering-ai-red-teaming",
    "title": "Prompt Engineering & AI Red Teaming",
    "url": "https://ai.engineer/talks/prompt-engineering-ai-red-teaming",
    "videoId": "_BRhRh7mOX0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 7265000,
    "speakers": [
      {
        "slug": "sander-schulhoff",
        "name": "Sander Schulhoff"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Sander Schulhoff, CEO of Learn Prompting and [REDACTED:username], leads an interactive workshop on research[REDACTED:generic_id]backed prompt engineering and adversarial AI security. He discusses the limits of role prompting, when fine[REDACTED:generic_id]tuning or retrieval[REDACTED:generic_id]based examples are more appropriate, findings associated with The Prompt Report, and audience questions about benchmarks, reinforcement learning, and prompt automation. The workshop then explores prompt injection, jailbreaks, agent[REDACTED:generic_id]security challenges, and a live [REDACTED:username] competition."
  },
  {
    "slug": "prompt-engineering-is-dead",
    "title": "Prompt Engineering is Dead",
    "url": "https://ai.engineer/talks/prompt-engineering-is-dead",
    "videoId": "jvKf6zXrNO4",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 859000,
    "speakers": [
      {
        "slug": "nir-gazit",
        "name": "Nir Gazit",
        "organization": "Traceloop"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Traceloop co-founder Nir Gazit argues that manually tuning prompts should give way to test-driven, automated optimization. Using a documentation chatbot built with a Chroma-backed RAG pipeline and OpenAI, he outlines how to create evaluation datasets, score responses with a ground-truth-based LLM judge, check answers against expected facts, and have an agent iteratively regenerate improved prompts. He closes by pointing attendees to Traceloop's public auto-prompting demo repository."
  },
  {
    "slug": "prompt-engineering-tactics",
    "title": "Prompt Engineering Tactics",
    "url": "https://ai.engineer/talks/prompt-engineering-tactics",
    "videoId": "7AYUCAuFYeA",
    "event": "AI Engineer Summit 2023",
    "durationMs": 312000,
    "speakers": [
      {
        "slug": "dan-cleary",
        "name": "Dan Cleary",
        "organization": "PromptHub"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "PromptHub co-founder Dan Cleary explains why nondeterministic LLM outputs and hallucinations threaten product quality, then presents three practical prompting tactics: multi-persona collaboration, grounding requests with an explicit source using an according-to clause, and EmotionPrompt, which appends emotional stimuli. He connects the final technique to published benchmark improvements and notes that PromptHub offers reusable templates."
  },
  {
    "slug": "prototyping-as-leadership-how-a-cto-ships-with-ai-agents",
    "title": "Prototyping as Leadership: How a CTO Ships with AI Agents",
    "url": "https://ai.engineer/talks/prototyping-as-leadership-how-a-cto-ships-with-ai-agents",
    "videoId": "bdHaOXZOhcM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1098000,
    "speakers": [
      {
        "slug": "hursh-agrawal",
        "name": "Hursh Agrawal",
        "organization": "The Browser Company"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "The Browser Company co-founder and CTO Hursh Agrawal argues that autonomous AI agents make hands-on prototyping a practical responsibility for engineering leaders with fragmented schedules. He outlines an overnight workflow: gather business and repository context from tools including Slack, Jira, Confluence, and Notion, launch an agent around 5 p.m., and review implementation, pull requests, CI results, and code-review fixes the next morning. He also applies the approach to evaluation-driven hill climbing and agent-directed model training using synthetic data and frontier-model ensembles."
  },
  {
    "slug": "pydantic-is-still-all-you-need",
    "title": "Pydantic is STILL all you need",
    "url": "https://ai.engineer/talks/pydantic-is-still-all-you-need",
    "videoId": "pZ4DIH2BVqg",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 921000,
    "speakers": [
      {
        "slug": "jason-liu",
        "name": "Jason Liu"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Jason Liu argues that Pydantic models and Instructor-style response schemas provide a more reliable, ergonomic foundation for LLM applications than parsing unstructured JSON. He discusses multilingual implementations, integrations with multiple model providers, typed response models, validated streaming, and field- and model-level validators, illustrating their usefulness for RAG, extraction, vision-based receipt processing, and identifying hallucinated URLs."
  },
  {
    "slug": "rag-and-the-mongodb-document-model",
    "title": "RAG and the MongoDB Document Model",
    "url": "https://ai.engineer/talks/rag-and-the-mongodb-document-model",
    "videoId": "2Ey275TX4ZU",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 793000,
    "speakers": [
      {
        "slug": "ben-flast",
        "name": "Ben Flast",
        "organization": "MongoDB"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "MongoDB's Ben Flast explains how retrieval-augmented generation supplies external context to large language models and how MongoDB's JSON/BSON document model combines application data with vector retrieval. He describes Atlas Vector Search, HNSW approximate-nearest-neighbor indexes, the $vectorSearch aggregation stage, similarity settings and numCandidates tuning, then highlights MongoDB-backed semantic caching with LangChain."
  },
  {
    "slug": "rag-at-scale-production-ready-genai-apps-with-azure-ai-search",
    "title": "RAG at scale: production-ready GenAI apps with Azure AI Search",
    "url": "https://ai.engineer/talks/rag-at-scale-production-ready-genai-apps-with-azure-ai-search",
    "videoId": "_2tZaDs-w5s",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1313000,
    "speakers": [
      {
        "slug": "pablo-castro",
        "name": "Pablo Castro",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Microsoft’s Pablo Castro explains how Azure AI Search supports production-scale retrieval-augmented generation, focusing on enterprise-ready retrieval infrastructure, vector search with HNSW and exhaustive nearest-neighbor methods, filtering, two-stage retrieval with L2 reranking, and continuously updated pipelines for chunking, vectorization, and indexing. Audience questions address nearest-neighbor result counts and keyword extraction from user prompts."
  },
  {
    "slug": "rag-evaluation-is-broken-here-s-why-and-how-to-fix-it",
    "title": "RAG Evaluation Is Broken! Here's Why (And How to Fix It)",
    "url": "https://ai.engineer/talks/rag-evaluation-is-broken-here-s-why-and-how-to-fix-it",
    "videoId": "Ywl4LsvHKzU",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 658000,
    "speakers": [
      {
        "slug": "yuval-belfer",
        "name": "Yuval Belfer",
        "organization": "AI21 Labs"
      },
      {
        "slug": "niv-granot",
        "name": "Niv Granot",
        "organization": "AI21 Labs"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "AI21 Labs presenters Yuval Belfer and Niv Granot argue that conventional RAG benchmarks overreward questions answerable from individual chunks while neglecting realistic aggregation across documents. Using financial examples and a 22-document FIFA World Cup corpus, they report baseline accuracies of 5% and 11% and describe a structured-RAG alternative that clusters documents, extracts and populates schemas during ingestion, and answers queries through SQL at inference. They also discuss normalization challenges such as changing country names and jointly hosted tournaments."
  },
  {
    "slug": "rag-for-vps-of-ai",
    "title": "RAG for VPs of AI",
    "url": "https://ai.engineer/talks/rag-for-vps-of-ai",
    "videoId": "KJsUHnwSvTY",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1611000,
    "speakers": [
      {
        "slug": "jerry-liu",
        "name": "Jerry Liu",
        "organization": "LlamaIndex"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "LlamaIndex co-founder and CEO Jerry Liu explains enterprise RAG architecture, including LLMs, embeddings, databases, document ingestion, and parsing, and argues that developer-built systems adapt better than packaged tools. Audience questions address LlamaCloud privacy, ETL skills, audio ingestion and semantic chunking, staging and production, regional availability, larger context windows, hallucinations, and agentic enterprise applications."
  },
  {
    "slug": "rag-in-2025-state-of-the-art-and-the-road-forward",
    "title": "RAG in 2025: State of the Art and the Road Forward",
    "url": "https://ai.engineer/talks/rag-in-2025-state-of-the-art-and-the-road-forward",
    "videoId": "W_CYk2ogcDI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1128000,
    "speakers": [
      {
        "slug": "tengyu-ma",
        "name": "Tengyu Ma",
        "organization": "MongoDB / Voyage AI"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Voyage AI co-founder Tengyu Ma explains why he favors retrieval-augmented generation over fine-tuning and long-context approaches for incorporating proprietary enterprise knowledge. He examines embedding models, vector databases, retrieval accuracy, Matryoshka learning and quantization, reranking, query and document enrichment, contextual retrieval, and domain-specific code embeddings. He concludes that increasingly capable retrieval models will absorb more of the hand-engineered techniques currently required for effective RAG."
  },
  {
    "slug": "rag-is-dead-right-kuba-rogut-redacted-username",
    "title": "RAG is dead, right?? — Kuba Rogut,",
    "url": "https://ai.engineer/talks/rag-is-dead-right-kuba-rogut-redacted-username",
    "videoId": "UM6sFg_jdlE",
    "event": "AI Engineer Europe 2026",
    "durationMs": 673000,
    "speakers": [],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "[REDACTED:username] engineer Kuba Rogut argues that RAG is evolving into agentic retrieval combining vector search, BM25, grep, glob, regex, and filters rather than disappearing. He contrasts Claude Code’s iterative filesystem search with Cursor’s pre-indexed semantic retrieval, explains secure reuse of codebase embeddings, cites Cursor’s measured 12.5% average answer-accuracy improvement, and frames embeddings as cached compute that can reduce runtime tokens, latency, and cost."
  },
  {
    "slug": "ralph-loops-build-dumb-ai-loops-that-ship",
    "title": "Ralph Loops: Build Dumb AI Loops That Ship",
    "url": "https://ai.engineer/talks/ralph-loops-build-dumb-ai-loops-that-ship",
    "videoId": "2TLXsxkz0zI",
    "event": "AI Engineer Europe 2026",
    "durationMs": 6506000,
    "speakers": [
      {
        "slug": "chris-parsons",
        "name": "Chris Parsons",
        "organization": "Cherrypick"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Cherrypick co-founder and CTO Chris Parsons leads a practical workshop on building simple Ralph Loops with AI coding agents. Using a public Pomodoro CLI repository, he demonstrates ticket-driven implementation, repeated Claude invocations, fresh session context, and agent skills, then discusses atomic commits, development methodologies, reviewing generated changes, evaluation, and when additional agent coordination becomes unnecessary complexity."
  },
  {
    "slug": "reachy-mini-the-300-open-source-robot-you-can-actually-hack-andres-marafioti-hugging-face",
    "title": "Reachy Mini: the $300 open source robot you can actually hack — Andres Marafioti, Hugging Face",
    "url": "https://ai.engineer/talks/reachy-mini-the-300-open-source-robot-you-can-actually-hack-andres-marafioti-hugging-face",
    "videoId": "0jeZfjJMfmo",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1276000,
    "speakers": [
      {
        "slug": "andres-marafioti",
        "name": "Andres Marafioti",
        "organization": "Hugging Face"
      }
    ],
    "topics": [
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Hugging Face's Andres Marafioti introduces Reachy Mini, an open-source conversational robot aimed at developers, researchers, and students. He describes its $300 and $450 configurations, including a wireless Raspberry Pi-powered model, and argues that mature speech models make voice a natural interface for embodied AI. Demonstrations cover conversational and multimodal robot interactions, followed by Qwen3-TTS optimization: the presented FasterQwenTTS system reports first-token latency below 200 milliseconds and roughly four-times-real-time synthesis."
  },
  {
    "slug": "real-ai-agents-need-planning-not-just-prompting",
    "title": "Real AI Agents Need Planning, Not Just Prompting",
    "url": "https://ai.engineer/talks/real-ai-agents-need-planning-not-just-prompting",
    "videoId": "Th5e4h-oVmc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 478000,
    "speakers": [
      {
        "slug": "yuval-belfer",
        "name": "Yuval Belfer",
        "organization": "AI21 Labs"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Yuval Belfer of AI21 Labs argues that complex instruction following requires agents that plan and adapt rather than rely solely on larger prompts. He contrasts routing, tool use, MCP, and ReAct with dynamic planning and replanning, then describes AI21 Maestro's planner and execution engine, including Best-of-N generation and decisions informed by expected cost, latency, and success probability. He reports improvements on IFEval and an internal requirement-satisfaction dataset while recommending the simplest architecture that solves the task."
  },
  {
    "slug": "real-roi-lessons-from-enterprises-that-have-already-succeeded-with-llms-redacted-username-scale",
    "title": "Real ROI: Lessons from Enterprises that Have already succeeded with LLMs Scale",
    "url": "https://ai.engineer/talks/real-roi-lessons-from-enterprises-that-have-already-succeeded-with-llms-redacted-username-scale",
    "videoId": "khcMErjUB8k",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1201000,
    "speakers": [
      {
        "slug": "raza-habib",
        "name": "Raza Habib",
        "organization": "Humanloop"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Humanloop cofounder and CEO Raza Habib explains how enterprises generate revenue and reduce costs with production LLM applications. He simplifies application architecture into models, prompts, data selection or retrieval, and optional function calling; cites GitHub Copilot and customer examples including Duolingo, Filevine, and Ironclad; and argues that successful teams combine generalist product engineers with domain experts who shape prompts and evaluations. He emphasizes evaluation-first development, manual annotation, and avoiding unnecessary orchestration complexity."
  },
  {
    "slug": "real-time-experiments-with-an-ai-co-scientist",
    "title": "Real-time Experiments with an AI Co-Scientist",
    "url": "https://ai.engineer/talks/real-time-experiments-with-an-ai-co-scientist",
    "videoId": "wNH3q9pqn0U",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1122000,
    "speakers": [
      {
        "slug": "stefania-druga",
        "name": "Stefania Druga"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Stefania Druga demonstrates a real-time AI co-scientist that connects micro:bit/Jacdac sensors, cameras, and laboratory instruments to Gemini-based reasoning. She relates the system to Google's Gemini 2.0 multi-agent AI co-scientist, explains a React/WebUSB sensor-to-model architecture, and discusses autonomous visual tracking, crystal-growth analysis, open-source laboratory automation, and science education."
  },
  {
    "slug": "real-world-development-with-github-copilot-and-vs-code-harald-kirschner-christopher-harrison",
    "title": "Real World Development with GitHub Copilot and VS Code — Harald Kirschner, Christopher Harrison",
    "url": "https://ai.engineer/talks/real-world-development-with-github-copilot-and-vs-code-harald-kirschner-christopher-harrison",
    "videoId": "eOxOzcw70f0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 4773000,
    "speakers": [
      {
        "slug": "harald-kirschner",
        "name": "Harald Kirschner",
        "organization": "Microsoft"
      },
      {
        "slug": "christopher-harrison",
        "name": "Christopher Harrison",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "An interactive, largely improvised GitHub Copilot and VS Code workshop explores outcome-first vibe coding, workspace-scoped agent settings and approvals, custom instructions, and reusable testing workflows. Audience questions prompt comparisons with Cursor and discussion of Claude, Three.js, and Perplexity, before demonstrations turn to MCP, GistPad-backed knowledge and reusable prompts, and server-to-client language-model sampling."
  },
  {
    "slug": "real-world-mcps-in-github-copilot-agent-mode",
    "title": "Real world MCPs in GitHub Copilot Agent Mode",
    "url": "https://ai.engineer/talks/real-world-mcps-in-github-copilot-agent-mode",
    "videoId": "RkVILz06y08",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 867000,
    "speakers": [
      {
        "slug": "jon-peck",
        "name": "Jon Peck",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Jon Peck demonstrates how GitHub Copilot agent mode uses Model Context Protocol servers to complete practical software-development tasks. He contrasts agent workflows with code completion and chat, shows how a detailed README can guide implementation, explains local and remote MCP connections, and demonstrates configuring a PostgreSQL MCP server in VS Code to retrieve database information and generate test fixtures. He also highlights repository-wide Copilot instructions and addresses an audience question about assigning issues to a coding agent."
  },
  {
    "slug": "realtime-conversational-video-with-pipecat-and-tavus-chad-bailey-and-brian-johnson-daily-tavus",
    "title": "Realtime Conversational Video with Pipecat and Tavus — Chad Bailey and Brian Johnson, Daily & Tavus",
    "url": "https://ai.engineer/talks/realtime-conversational-video-with-pipecat-and-tavus-chad-bailey-and-brian-johnson-daily-tavus",
    "videoId": "ujt0da9Z29Q",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1126000,
    "speakers": [
      {
        "slug": "chad-bailey",
        "name": "Chad Bailey",
        "organization": "Daily"
      },
      {
        "slug": "brian-johnson",
        "name": "Brian Johnson",
        "organization": "Tavus"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Chad Bailey of Daily and Brian Johnson of Tavus explain how to build real-time conversational video agents by combining voice-model pipelines, Pipecat orchestration, and Tavus digital-human video. They describe cascaded speech-to-text, LLM, and text-to-speech systems; Tavus's Sparrow-0 and Raven-0 models; Pipecat's frames, processors, streaming pipelines, and voice activity detection; WebRTC, WebSocket, and Twilio transports; audiovisual synchronization; and deployment with Pipecat Cloud."
  },
  {
    "slug": "realtime-multiplayer-automation-and-you",
    "title": "Realtime multiplayer, automation, and you!",
    "url": "https://ai.engineer/talks/realtime-multiplayer-automation-and-you",
    "videoId": "iQ5xldZ9StU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1301000,
    "speakers": [
      {
        "slug": "idan-gazit",
        "name": "Idan Gazit",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "GitHub Next leader Idan Gazit presents two complementary approaches to agent-assisted software development: Agentic Workflows, which compile natural-language Markdown instructions into governed GitHub Actions automation, and Ace, a research prototype for real-time collaboration between teammates and coding agents. An Astro dependency-upgrade example illustrates practical automation, while YAML-defined permissions, network restrictions, mediated secrets, and auditable writes address agent safety. Additional use cases include issue triage, CI diagnosis, repository reporting, and preserving shared team context."
  },
  {
    "slug": "recsys-keynote-improving-recommendation-systems-search-in-the-age-of-llms",
    "title": "Recsys Keynote: Improving Recommendation Systems & Search in the Age of LLMs",
    "url": "https://ai.engineer/talks/recsys-keynote-improving-recommendation-systems-search-in-the-age-of-llms",
    "videoId": "2vlCqD6igVA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1254000,
    "speakers": [
      {
        "slug": "eugene-yan",
        "name": "Eugene Yan",
        "organization": "Amazon"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Eugene Yan outlines three ways language-model techniques are reshaping recommendation and search: semantic item IDs that incorporate multimodal content and reduce cold-start problems; LLM-generated metadata, labels, and synthetic training data; and unified foundation-style models spanning multiple tasks. Examples include Kuaishou’s two-tower architecture using ResNet, BERT, VGGish, and clustered semantic IDs, plus search and discovery challenges at Spotify and Indeed."
  },
  {
    "slug": "recursive-coding-agents",
    "title": "Recursive Coding Agents",
    "url": "https://ai.engineer/talks/recursive-coding-agents",
    "videoId": "3hXJI2q0Jz8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1428000,
    "speakers": [
      {
        "slug": "raymond-weitekamp",
        "name": "Raymond Weitekamp",
        "organization": "OpenProse"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Raymond Weitekamp of OpenProse argues that dependable coding agents require outcome-oriented orchestration rather than simply more capable models. He explains Recursive Language Models as systems that externalize context into a REPL, manipulate it symbolically, and delegate work recursively; discusses long-context processing, memory, DSPy, Oolong, LongCoT, and claimed small-model reasoning results; and contrasts Claude Code dynamic workflows with OpenProse approaches to recursive agent execution."
  },
  {
    "slug": "recursive-model-improvement",
    "title": "Recursive Model Improvement",
    "url": "https://ai.engineer/talks/recursive-model-improvement",
    "videoId": "q4Tr-DknG2M",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1232000,
    "speakers": [
      {
        "slug": "lee-robinson",
        "name": "Lee Robinson",
        "organization": "Cursor"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Cursor's Lee Robinson explains recursive model improvement as a flywheel combining real-world user feedback, private evaluations, reinforcement-learning environments, and increasingly capable coding models. He discusses Composer 2.5, the earlier Kimi-based training approach, Cursor's held-out coding benchmark, agent self-coaching, compute allocation across training stages, Colossus and TeraFab infrastructure, and automation of machine-learning research."
  },
  {
    "slug": "reinforcement-learning-without-verifiable-rewards-will-brown-prime-intellect",
    "title": "Reinforcement Learning without Verifiable Rewards — Will Brown, Prime Intellect",
    "url": "https://ai.engineer/talks/reinforcement-learning-without-verifiable-rewards-will-brown-prime-intellect",
    "videoId": "AQv3qRCG6Gw",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1167000,
    "speakers": [
      {
        "slug": "will-brown",
        "name": "Will Brown",
        "organization": "Prime Intellect"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Prime Intellect applied-research lead Will Brown explains how to extend reinforcement learning beyond tasks with easily verifiable rewards. He reviews policy-gradient methods and Prime Intellect's PRIME-RL and Lab infrastructure, then discusses reusable environments, reward hacking, distributional uncertainty, LLM judges, search-derived rubrics, supervision from real code changes, simulated tools, and difficulty calibration for practical agent training."
  },
  {
    "slug": "remobi-app-don-t-change-your-terminal-workflow-for-mobile",
    "title": "remobi.app: Don't change your terminal workflow for mobile",
    "url": "https://ai.engineer/talks/remobi-app-don-t-change-your-terminal-workflow-for-mobile",
    "videoId": "5192csoTkVo",
    "event": "AI Engineer Europe 2026",
    "durationMs": 561000,
    "speakers": [
      {
        "slug": "connor-adams",
        "name": "Connor Adams"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Connor Adams introduces Remobi, an open-source mobile terminal interface for monitoring and controlling coding agents without abandoning existing tmux and remote-development workflows. He contrasts it with Claude Code-specific mobile clients, manual session handoffs, and conventional SSH apps; demonstrates customizable terminal controls; and describes its iOS- and Android-compatible progressive web app, npm installation, and GitHub repository. The presentation concludes with an audience question."
  },
  {
    "slug": "remote-mcps-what-we-learned-from-shipping-john-welsh-anthropic",
    "title": "Remote MCPs: What we learned from shipping — John Welsh, Anthropic",
    "url": "https://ai.engineer/talks/remote-mcps-what-we-learned-from-shipping-john-welsh-anthropic",
    "videoId": "0NHCyq8bBcM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 912000,
    "speakers": [
      {
        "slug": "john-welsh",
        "name": "John Welsh",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Anthropic engineer John Welsh explains lessons from implementing remote MCP clients and standardizing internal and external model-tool integrations on Model Context Protocol. He distinguishes JSON-RPC message semantics from Streamable HTTP transport and OAuth 2.1, discusses built-in sampling primitives, and recommends centralizing authentication, user credentials, and external connectivity at a shared gateway while allowing organization-specific internal transports."
  },
  {
    "slug": "replacing-12k-loc-with-a-200-loc-skill",
    "title": "Replacing 12K LoC with a 200 LoC Skill",
    "url": "https://ai.engineer/talks/replacing-12k-loc-with-a-200-loc-skill",
    "videoId": "WE_Gnowy3uw",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1162000,
    "speakers": [
      {
        "slug": "david-gomes",
        "name": "David Gomes",
        "organization": "Cursor"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "David Gomes explains how Cursor replaced its heavyweight Git-worktree implementation with a lightweight Markdown skill, slash commands, and subagents supporting isolated parallel coding, Best-of-N model comparisons, and multi-repository workflows. He contrasts reduced implementation complexity and greater flexibility with weaker isolation guarantees, reduced discoverability, and inconsistent agent behavior, then outlines improvements using evaluations, reinforcement learning, Cursor's Composer model, and system reminders."
  },
  {
    "slug": "research-to-reality-bringing-frontier-ml-research-to-production",
    "title": "Research to Reality: Bringing frontier ML research to production",
    "url": "https://ai.engineer/talks/research-to-reality-bringing-frontier-ml-research-to-production",
    "videoId": "OXMMN-XbxwA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 897000,
    "speakers": [
      {
        "slug": "vaidas-razgaitis",
        "name": "Vaidas Razgaitis",
        "organization": "Higharc"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Higharc research engineer Vaidas Razgaitis explains how his team turns frontier machine-learning research into production homebuilding features. He describes improving research legibility with technical design documents covering domain context, business goals, data, persistence, and architecture; organizing decoupled Python services in a monorepo behind a Docker-networked gateway; exposing layered services with FastAPI; and supporting delivery with Poetry or uv, GitHub Actions, automated tests, and dependency-aware project decomposition."
  },
  {
    "slug": "research-to-reality-with-google-deepmind",
    "title": "Research to Reality with Google DeepMind",
    "url": "https://ai.engineer/talks/research-to-reality-with-google-deepmind",
    "videoId": "1P1hJ36rxM0",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1226000,
    "speakers": [
      {
        "slug": "benoit-schillings",
        "name": "Benoit Schillings",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Google DeepMind's Benoit Schillings traces his team's Gemini research and the early Pitchfork coding project, arguing that AI makes generating code comparatively easy while software architecture, understanding large codebases, and planning remain difficult. He describes dwindling human-written training data and proposes AlphaZero-style self-play and reinforcement learning as routes toward more capable coding models and novel algorithmic discovery."
  },
  {
    "slug": "respect-the-process",
    "title": "Respect The Process",
    "url": "https://ai.engineer/talks/respect-the-process",
    "videoId": "CLttOU7n6sI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1003000,
    "speakers": [
      {
        "slug": "andrew-dumit",
        "name": "Andrew Dumit",
        "organization": "Watershed Technology Inc."
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Watershed AI engineer Andrew Dumit explains why sustainability calculations require evaluating an agent’s reasoning process, not just its final answer. He describes replacing narrowly specified graph-manipulation tools with coding agents that can explore and edit complex emissions graphs, while restricting consequential changes to a typed TypeScript SDK and organization-controlled deterministic execution. This architecture validates graph edits, produces inspectable artifacts for nonprogrammer experts, and is reinforced through SDK documentation, few-shot examples, improved tool ergonomics, and prompt engineering."
  },
  {
    "slug": "rethinking-environments-for-long-horizon-work",
    "title": "Rethinking Environments for Long Horizon Work",
    "url": "https://ai.engineer/talks/rethinking-environments-for-long-horizon-work",
    "videoId": "2aS7aKoXn64",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1275000,
    "speakers": [
      {
        "slug": "rayan-garg",
        "name": "Rayan Garg",
        "organization": "Theta Software"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Theta Software's Rayan Garg and an unnamed co-presenter examine how to define and evaluate long-horizon AI-agent work. They contrast METR's human-time success thresholds with token- and trajectory-based measures, explain why human effort can misrepresent model difficulty, and discuss multi-tool reinforcement-learning environments. The talk also describes judge or critic models that assess final environment state and agent trajectories, queryable execution traces, benchmark saturation, and mean@5 reporting."
  },
  {
    "slug": "rethinking-team-building-how-a-30-person-startup-serves-50-million-users-grant-lee-gamma",
    "title": "Rethinking Team Building: How a 30-Person Startup Serves 50 Million Users — Grant Lee, Gamma",
    "url": "https://ai.engineer/talks/rethinking-team-building-how-a-30-person-startup-serves-50-million-users-grant-lee-gamma",
    "videoId": "q8zoXAbmJdI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1086000,
    "speakers": [
      {
        "slug": "grant-lee",
        "name": "Grant Lee",
        "organization": "Gamma"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Gamma co-founder and CEO Grant Lee argues that AI startups should rethink organizational design instead of reflexively expanding management layers. He advocates hiring adaptable multidisciplinary generalists, empowering hands-on player-coach leaders, and reinforcing shared culture through a living culture deck, product dogfooding, and company-wide show-and-tells. Audience questions explore infrastructure choices, integrating specialist hires, evaluating high agency, and using work trials."
  },
  {
    "slug": "retrieval-augmented-generation-in-the-wild",
    "title": "Retrieval Augmented Generation in the Wild",
    "url": "https://ai.engineer/talks/retrieval-augmented-generation-in-the-wild",
    "videoId": "FDEmbYPgG-s",
    "event": "AI Engineer Summit 2023",
    "durationMs": 740000,
    "speakers": [
      {
        "slug": "anton-troynikov",
        "name": "Anton Troynikov",
        "organization": "Chroma"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Chroma co-founder Anton Troynikov argues that production retrieval-augmented generation requires more than basic vector search. He describes feedback-aware and self-updating agent memory, Voyager’s retrievable Minecraft skills, context distractors, use-case-specific evaluation, chunking, hierarchical retrieval, re-ranking, irrelevant-result detection, and multimodal data infrastructure."
  },
  {
    "slug": "revenue-engineering-how-to-price-and-reprice-your-ai-product",
    "title": "Revenue Engineering: How to Price (and Reprice) Your AI Product",
    "url": "https://ai.engineer/talks/revenue-engineering-how-to-price-and-reprice-your-ai-product",
    "videoId": "1C3sZbaxOmw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 939000,
    "speakers": [
      {
        "slug": "kshitij-grover",
        "name": "Kshitij Grover",
        "organization": "Orb"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Orb co-founder and CTO Kshitij Grover explains how AI companies should design pricing around customer value, purchasing behavior, variable infrastructure costs, and predictable budgets. Using Replit Agent and Cloudflare Workers as examples, he contrasts self-service and enterprise sales, shows how CPU-based infrastructure pricing can align with AI workloads, and discusses usage-based incentives, ongoing repricing, threshold billing, fraud, and prepaid models."
  },
  {
    "slug": "reverse-conway-s-law-and-genai-how-agents-will-take-over-the-organisation",
    "title": "Reverse Conway's law and GenAI: How agents will take over the organisation",
    "url": "https://ai.engineer/talks/reverse-conway-s-law-and-genai-how-agents-will-take-over-the-organisation",
    "videoId": "FpJ9dPe1qYQ",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1692000,
    "speakers": [
      {
        "slug": "patrick-debois",
        "name": "Patrick Debois",
        "organization": "Jedi"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Patrick Debois applies Reverse Conway's Law to generative AI, examining how agents could reshape organizational structures, engineering roles, and team composition. He traces a progression from copilots and coding contributors to domain-specific teammates and potential managers, discussing task unbundling, paid AI-training work, cross-domain responsibilities, junior-developer learning, and the need for human verification, judgment, and failure awareness."
  },
  {
    "slug": "reverse-engineering-a-viking-voip-phone-protocol-with-claude-code-boris-starkov-elevenlabs",
    "title": "Reverse engineering a Viking VoIP phone protocol with Claude Code — Boris Starkov, ElevenLabs",
    "url": "https://ai.engineer/talks/reverse-engineering-a-viking-voip-phone-protocol-with-claude-code-boris-starkov-elevenlabs",
    "videoId": "V-L0INGTEOg",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1211000,
    "speakers": [
      {
        "slug": "boris-starkov",
        "name": "Boris Starkov",
        "organization": "ElevenLabs"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "ElevenLabs engineer Boris Starkov explains how he used Claude Code to reverse-engineer an undocumented Viking VoIP phone whose configuration software required Windows XP. He describes brute-forcing two-letter device commands, diagnosing settings that disappeared after reboot, and validating discoveries through closed-loop experimentation. The resulting architecture routes the phone through a Twilio SIP trunk into an ElevenLabs conversational agent, powering an interactive Michael Caine-voiced phone-booth demonstration."
  },
  {
    "slug": "reviewdebt-a-practical-framework-for-scoring-every-pull-request-sachin-gupta-ebay",
    "title": "ReviewDebt: a practical framework for scoring every pull request — Sachin Gupta, eBay",
    "url": "https://ai.engineer/talks/reviewdebt-a-practical-framework-for-scoring-every-pull-request-sachin-gupta-ebay",
    "videoId": "TJPInBjhE4Q",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1500000,
    "speakers": [
      {
        "slug": "sachin-gupta",
        "name": "Sachin Gupta",
        "organization": "eBay"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Sachin Gupta introduces ReviewDebt, a deterministic zero-to-100 framework for estimating the human review burden created by pull requests, particularly AI-generated changes. Drawing on GitHub and Faros AI research, he critiques throughput-only productivity metrics, examines diff sprawl, cross-file coupling, weak test evidence and reviewer fatigue, and illustrates practical scoring, escalation thresholds and a scan of 524 pull requests across public repositories."
  },
  {
    "slug": "rewiring-the-state",
    "title": "Rewiring the State",
    "url": "https://ai.engineer/talks/rewiring-the-state",
    "videoId": "ObNKGf9YR0g",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1698000,
    "speakers": [
      {
        "slug": "eoin-mulgrew",
        "name": "Eoin Mulgrew"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Eoin Mulgrew explains how the Number 10 Data Science team is building a fellowship and forward-deployed AI engineering capability to improve UK public services. He discusses recruiting external technical talent, practical government delivery constraints, AI-assisted analysis of the UK statute book, operational delivery red-teaming, and the Inspect framework for evaluating autonomous AI agents. Audience questions address government deployment, AI tutoring, and international collaboration."
  },
  {
    "slug": "rise-of-the-ai-architect-clay-bavor-and-alessio-fanelli",
    "title": "Rise of the AI Architect — Clay Bavor and Alessio Fanelli",
    "url": "https://ai.engineer/talks/rise-of-the-ai-architect-clay-bavor-and-alessio-fanelli",
    "videoId": "C3geUfBR2js",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1135000,
    "speakers": [
      {
        "slug": "clay-bavor",
        "name": "Clay Bavor",
        "organization": "Sierra"
      },
      {
        "slug": "alessio-fanelli",
        "name": "Alessio Fanelli",
        "organization": "Latent.Space"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "In a fireside chat with Alessio Fanelli, Sierra cofounder Clay Bavor describes the emerging AI architect as the person responsible for shaping, coaching, and improving a company’s customer-facing AI agent. He compares the role to an early webmaster and argues that it combines technical awareness, brand and experience design, and business judgment. Examples involving ADT, SiriusXM’s Harmony, and Chubbies illustrate branded agents, while customer-experience teams and ongoing conversation review demonstrate how organizations can develop and continuously refine these systems."
  },
  {
    "slug": "rl-environments-at-scale",
    "title": "RL Environments at Scale",
    "url": "https://ai.engineer/talks/rl-environments-at-scale",
    "videoId": "_IzZWeuTx7I",
    "event": "AI Engineer Code 2025",
    "durationMs": 1110000,
    "speakers": [
      {
        "slug": "will-brown",
        "name": "Will Brown",
        "organization": "Prime Intellect"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Will Brown presents Prime Intellect's approach to scaling reinforcement-learning environments and broadening access to AI research. He describes environments as reusable combinations of tasks, agent harnesses, and rewards that support evaluations, synthetic data, supervised fine-tuning, distillation, and reinforcement learning. The talk highlights product-specific model training, the Environments Hub, the Verifiers library, and PRIME-RL, including a reported post-training run involving a 100B-plus-parameter model and 500 GPUs."
  },
  {
    "slug": "rl-for-autonomous-coding-aakanksha-chowdhery-reflection-ai",
    "title": "RL for Autonomous Coding — Aakanksha Chowdhery, Reflection AI",
    "url": "https://ai.engineer/talks/rl-for-autonomous-coding-aakanksha-chowdhery-reflection-ai",
    "videoId": "QluDzKVfp6A",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1167000,
    "speakers": [
      {
        "slug": "aakanksha-chowdhery",
        "name": "Aakanksha Chowdhery",
        "organization": "Reflection AI"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Aakanksha Chowdhery of Reflection AI traces the progression from language-model scaling and emergent reasoning in PaLM to inference-time sampling, verifiable code generation, and reinforcement learning. She explains how SWE-bench Verified, unit tests, execution feedback, compiler-based verification, and DeepSeekMath’s GRPO inform reward design for autonomous coding agents, then discusses Reflection’s real-world software-engineering ambitions and audience questions about simulation versus experience."
  },
  {
    "slug": "rlm-recursive-language-models-for-large-codebases",
    "title": "RLM: Recursive Language Models for Large Codebases",
    "url": "https://ai.engineer/talks/rlm-recursive-language-models-for-large-codebases",
    "videoId": "8oyalrfwgjw",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1047000,
    "speakers": [
      {
        "slug": "shashi-jagtap",
        "name": "Shashi Jagtap",
        "organization": "Superagentic AI"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Superagentic AI founder Shashi Jagtap explains why coding agents struggle with large repositories and demonstrates Recursive Language Models as an alternative: keep repository context in a programmable REPL, inspect it through generated code, return bounded observations, and recursively delegate focused questions to other language-model calls. He introduces the open-source RLM Code research playground, compares it with author-maintained RLM and DSPy.RLM implementations, and demonstrates an inspectable, budget-controlled execution workflow."
  },
  {
    "slug": "road-to-5-million-tokens-breaking-barriers-in-long-context-training-max-ryabinin-together-ai",
    "title": "Road to 5 Million Tokens: Breaking Barriers in Long Context Training — Max Ryabinin, Together AI",
    "url": "https://ai.engineer/talks/road-to-5-million-tokens-breaking-barriers-in-long-context-training-max-ryabinin-together-ai",
    "videoId": "TUnPNY4E2fw",
    "event": "AI Engineer Europe 2026",
    "durationMs": 950000,
    "speakers": [
      {
        "slug": "max-ryabinin",
        "name": "Max Ryabinin",
        "organization": "Together AI"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Together AI researcher Max Ryabinin explains how GPU memory limits constrain multi-million-token transformer training and builds a solution from fully sharded data parallelism, DeepSpeed-Ulysses context parallelism, CPU activation offloading, and chunked computations. He then introduces UPipe, the Untied Ulysses technique that processes attention heads in smaller groups to reduce activation memory and enable five-million-token contexts on an eight-H100 node, before discussing profiling and taking audience questions."
  },
  {
    "slug": "robotics-why-now",
    "title": "Robotics: why now?",
    "url": "https://ai.engineer/talks/robotics-why-now",
    "videoId": "cGLa8DsOYdk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1087000,
    "speakers": [
      {
        "slug": "quan-vuong",
        "name": "Quan Vuong",
        "organization": "Physical Intelligence"
      },
      {
        "slug": "jost-tobias-springenberg",
        "name": "Jost Tobias Springenberg",
        "organization": "Physical Intelligence"
      }
    ],
    "topics": [
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "Quan Vuong and Jost Tobias Springenberg explain why advances in general AI and vision-language-action models are accelerating general-purpose robotics. They describe adapting pretrained vision-language backbones for high-frequency robot control, building a teleoperation-driven data engine, and using π0-style action experts and flow matching to execute continuous actions and decompose complex instructions. Demonstrations highlight autonomous manipulation, household cleaning in previously unseen environments, and end-to-end coffee preparation, alongside unresolved scientific, engineering, and operational challenges."
  },
  {
    "slug": "running-a-chess-youtube-channel-entirely-by-ai-stephan-steinfurt-tng-technology-consulting",
    "title": "Running a Chess YouTube Channel Entirely by AI — Stephan Steinfurt, TNG Technology Consulting",
    "url": "https://ai.engineer/talks/running-a-chess-youtube-channel-entirely-by-ai-stephan-steinfurt-tng-technology-consulting",
    "videoId": "BqZrTdgBaPw",
    "event": "AI Engineer Europe 2026",
    "durationMs": 991000,
    "speakers": [
      {
        "slug": "stephan-steinfurt",
        "name": "Stephan Steinfurt",
        "organization": "TNG Technology Consulting"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "Stephan Steinfurt of TNG Technology Consulting demonstrates an autonomous chess YouTube workflow that downloads Lichess games nightly, analyzes positions with an LLM agent and specialized chess tools, and automatically publishes narrated puzzle videos. A generated example explains a rook sacrifice leading to a decisive knight fork. He discusses Gemini 3.1 Pro, earlier experiments with Grok 4 and OpenAI models, the checks-captures-threats analysis tool, and audience questions about video quality and cost."
  },
  {
    "slug": "running-ai-application-in-minutes-quick-start-with-ai-templates",
    "title": "Running AI Application in Minutes: Quick Start with AI Templates",
    "url": "https://ai.engineer/talks/running-ai-application-in-minutes-quick-start-with-ai-templates",
    "videoId": "1t-9-s1brcg",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 5342000,
    "speakers": [
      {
        "slug": "pamela-fox",
        "name": "Pamela Fox",
        "organization": "Microsoft"
      },
      {
        "slug": "harald-kirschner",
        "name": "Harald Kirschner",
        "organization": "Microsoft"
      },
      {
        "slug": "gabriela-de-queiroz",
        "name": "Gabriela de Queiroz",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Gabriela de Queiroz, Pamela Fox, and Harald Kirschner present a hands-on Microsoft workshop on rapidly building and deploying generative AI applications with AI templates, GitHub Codespaces, Azure OpenAI, and Azure Developer CLI. Demonstrations cover startup resources, workshop proxy configuration, local chat-app development, deployment settings, retrieval-augmented generation over PostgreSQL and Azure AI Search, document context limits, evaluation and hybrid retrieval, and GitHub Copilot inline chat."
  },
  {
    "slug": "running-ai-application-in-minutes-quick-start-with-ai-templates-ee4087",
    "title": "Running AI Application in Minutes: Quick Start with AI Templates",
    "url": "https://ai.engineer/talks/running-ai-application-in-minutes-quick-start-with-ai-templates-ee4087",
    "videoId": "a0WvaMRZ9hc",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 3800000,
    "speakers": [
      {
        "slug": "gabriela-de-queiroz",
        "name": "Gabriela de Queiroz",
        "organization": "Microsoft"
      },
      {
        "slug": "aishwarya-srinivasan",
        "name": "Aishwarya Srinivasan",
        "organization": "Microsoft"
      },
      {
        "slug": "pamela-fox",
        "name": "Pamela Fox",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Microsoft presenters introduce Microsoft for Startups, Founders Hub, and Build with AI, then guide a hands-on workshop using open-source AI application templates, GitHub Codespaces, and an Azure OpenAI proxy. The demonstration runs a Python Quart application, discusses Container Apps deployment, and evaluates search groundedness. Pamela Fox appears as an additional substantial presenter despite not appearing on the authoritative two-person scheduled roster."
  },
  {
    "slug": "running-gemma-4-on-device-40-tokens-s-on-iphone-with-mlx",
    "title": "Running Gemma 4 On-Device: 40 Tokens/s on iPhone with MLX",
    "url": "https://ai.engineer/talks/running-gemma-4-on-device-40-tokens-s-on-iphone-with-mlx",
    "videoId": "a2muGkT4WD4",
    "event": "AI Engineer Europe 2026",
    "durationMs": 651000,
    "speakers": [
      {
        "slug": "adrien-grondin",
        "name": "Adrien Grondin",
        "organization": "Locally AI"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Locally AI developer Adrien Grondin demonstrates running Google DeepMind’s Gemma 4 on an iPhone with Apple’s MLX framework, including a claimed 40-tokens-per-second demo. He explains native integration through mlx-swift-lm, selecting Hugging Face MLX Community models and quantized variants, and the broader MLX ecosystem for vision, audio, and video. The session closes with audience questions, including tool calling and model compatibility."
  },
  {
    "slug": "running-llms-locally-practical-llm-performance-on-dgx-spark-mozhgan-kabiri-chimeh-nvidia",
    "title": "Running LLMs locally: Practical LLM Performance on DGX Spark — Mozhgan Kabiri chimeh, NVIDIA",
    "url": "https://ai.engineer/talks/running-llms-locally-practical-llm-performance-on-dgx-spark-mozhgan-kabiri-chimeh-nvidia",
    "videoId": "c5-kx2bwoCk",
    "event": "AI Engineer Europe 2026",
    "durationMs": 616000,
    "speakers": [
      {
        "slug": "mozhgan-kabiri-chimeh",
        "name": "Mozhgan Kabiri Chimeh",
        "organization": "NVIDIA"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "NVIDIA Developer Relations Manager Mozhgan Kabiri Chimeh demonstrates reproducible local LLM benchmarking on DGX Spark, using vLLM, Qwen models, Docker isolation, warm-up runs, and GPU telemetry. She explains how the GB10 Grace Blackwell Superchip and 128 GB of unified memory support local development, compares throughput and time to first token across 1.5B–14B models, and reports that a 14B NVFP4 model reaches its first token 3.4 times faster than its unoptimized counterpart. The talk emphasizes quantization, memory-bandwidth constraints, privacy-sensitive workloads, and portability from local prototyping to cloud or data-center deployment."
  },
  {
    "slug": "rust-is-the-language-of-the-agi",
    "title": "Rust is the language of the AGI",
    "url": "https://ai.engineer/talks/rust-is-the-language-of-the-agi",
    "videoId": "bbq0b_FpYEY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1754000,
    "speakers": [
      {
        "slug": "michael-yuan",
        "name": "Michael Yuan"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Michael Yuan argues that Rust's strong type system and compiler diagnostics make it particularly suitable for AI-generated software because compilation provides a tight, verifiable feedback loop. He introduces the open-source RustCoder project, describes Linux Foundation mentorship support, demonstrates a Gaia-hosted coding workflow and Cursor integration with an MCP Compile and Fix tool, and discusses LlamaEdge and expanding Rust knowledge available to coding agents."
  },
  {
    "slug": "safety-and-security-for-code-executing-agents",
    "title": "Safety and security for code-executing agents",
    "url": "https://ai.engineer/talks/safety-and-security-for-code-executing-agents",
    "videoId": "w7IMuYsBNr8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 840000,
    "speakers": [
      {
        "slug": "fouad-matin",
        "name": "Fouad Matin",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "OpenAI engineer Fouad Matin explains how increasingly capable coding and multimodal agents expand security risks, especially remote code execution, prompt injection, and data exfiltration. Using Codex and its open-source CLI, he describes layered safeguards: isolated containers, restricted internet access, human review and approval policies, macOS Seatbelt sandboxing inspired by Chromium, and Rust-based Linux isolation using seccomp and Landlock."
  },
  {
    "slug": "scaffold-wisely",
    "title": "Scaffold Wisely",
    "url": "https://ai.engineer/talks/scaffold-wisely",
    "videoId": "-rsTkYgnNzM",
    "event": "AI Engineer Summit 2025",
    "durationMs": 992000,
    "speakers": [
      {
        "slug": "rahul-sengottuvelu",
        "name": "Rahul Sengottuvelu",
        "organization": "Ramp"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Ramp’s Rahul Sengottuvelu argues that AI agents should rely less on brittle, hand-engineered scaffolding and more on general-purpose models and scalable computation. After discussing Jsonformer and historical examples of general methods outperforming specialized systems, he compares three approaches to converting arbitrary financial CSV files: deterministic integrations, pipelines augmented with LLM calls, and an agent using a code interpreter plus executable verification. He closes with an experimental LLM-driven email client and considers how increasingly capable models could reshape software interfaces."
  },
  {
    "slug": "scaling-agents-on-kubernetes-with-acpx-and-acp",
    "title": "Scaling Agents on Kubernetes with acpx and ACP",
    "url": "https://ai.engineer/talks/scaling-agents-on-kubernetes-with-acpx-and-acp",
    "videoId": "VaS2h-dY1-4",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1140000,
    "speakers": [
      {
        "slug": "onur-solmaz",
        "name": "Onur Solmaz",
        "organization": "OpenClaw; TextCortex at the time of recording"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Onur Solmaz describes his OpenClaw maintainer work, earlier TextCortex coding-agent development, and the use of acpx and the Agent Client Protocol to orchestrate coding agents through structured sessions. He demonstrates parallel Codex workflows bound to Discord channels, contrasts ACP with Codex App Server integration, and discusses automated pull-request feedback, review loops, enterprise adoption, Kubernetes, and scalable multi-agent provisioning."
  },
  {
    "slug": "scaling-ai-agents-without-breaking-reliability",
    "title": "Scaling AI agents without breaking reliability",
    "url": "https://ai.engineer/talks/scaling-ai-agents-without-breaking-reliability",
    "videoId": "1izYWsokr9s",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 901000,
    "speakers": [
      {
        "slug": "preeti-somal",
        "name": "Preeti Somal",
        "organization": "Temporal"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Temporal engineering leader Preeti Somal explains why production AI agents should be treated as complex distributed systems requiring durable orchestration, persistent state, visibility, and automatic failure handling. She describes language-idiomatic SDKs, including Python, outlines production examples involving Dust and Gorgias, and shows how Temporal workflows, workers, open-source code, and existing CI/CD practices support reliable agent deployments."
  },
  {
    "slug": "scaling-ai-in-education-a-khanmigo-case-study",
    "title": "Scaling AI in Education: A Khanmigo case study",
    "url": "https://ai.engineer/talks/scaling-ai-in-education-a-khanmigo-case-study",
    "videoId": "3E7VAZaTG9M",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1359000,
    "speakers": [
      {
        "slug": "shawn-jansepar",
        "name": "Shawn Jansepar",
        "organization": "Khan Academy"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Khan Academy's Shawn Jansepar explains how Khanmigo scales personalized tutoring and teacher assistance through Socratic questioning, stronger mathematical accuracy, context-aware learning, and education-specific interfaces. He demonstrates iterative essay coaching and teacher visibility into student work, shares classroom feedback, describes rapid prototyping as an organizational practice, and discusses efficiently provisioning dedicated compute for fluctuating demand."
  },
  {
    "slug": "scaling-compute-on-context",
    "title": "Scaling Compute on Context",
    "url": "https://ai.engineer/talks/scaling-compute-on-context",
    "videoId": "WiqDvX6isc4",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1182000,
    "speakers": [
      {
        "slug": "jack-morris",
        "name": "Jack Morris",
        "organization": "Engram"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Engram researcher Jack Morris frames scaling compute on context as the challenge of giving pretrained language models deep, personalized knowledge beyond publicly available training data. He contrasts broad model knowledge with expertise in private information and rare skills such as AMD GPU kernel development, reviews scaling through data, compute, and model capacity, references METR task-horizon measurements, and discusses approaches including KV compaction and on-policy distillation for transferring knowledge from an unstructured corpus into a model."
  },
  {
    "slug": "scaling-enterprise-grade-rag-systems-lessons-from-the-legal-frontier",
    "title": "Scaling Enterprise-Grade RAG Systems: Lessons from the Legal Frontier",
    "url": "https://ai.engineer/talks/scaling-enterprise-grade-rag-systems-lessons-from-the-legal-frontier",
    "videoId": "W1MiZChnkfA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1000000,
    "speakers": [
      {
        "slug": "calvin-qi",
        "name": "Calvin Qi",
        "organization": "Harvey"
      },
      {
        "slug": "chang-she",
        "name": "Chang She",
        "organization": "LanceDB"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Calvin Qi of Harvey and Chang She of LanceDB explain how enterprise legal RAG must handle complex regulatory questions, domain-specific evaluation, strict confidentiality, customer isolation, retention requirements, and corpora containing tens of millions of documents. They describe balancing ingestion throughput and query latency, then introduce LanceDB’s multimodal lakehouse architecture for storing embeddings, text, and other AI data together."
  },
  {
    "slug": "scaling-github-for-your-agents-sam-morrow-github",
    "title": "Scaling GitHub for your Agents — Sam Morrow, GitHub",
    "url": "https://ai.engineer/talks/scaling-github-for-your-agents-sam-morrow-github",
    "videoId": "0n3MKk7r60w",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1235000,
    "speakers": [
      {
        "slug": "sam-morrow",
        "name": "Sam Morrow",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "GitHub MCP Server developer Sam Morrow explains challenges encountered while scaling GitHub’s remote MCP server, including excessive tool inventories, context-window consumption, default configuration behavior, semantic tool discovery, and grouped CRUD tools. He discusses practical tool evaluation, secure credential handling, OAuth 2.1 and PKCE, prompt-injection and private-data exfiltration risks, and experimental Insiders Mode features supporting human-in-the-loop workflows."
  },
  {
    "slug": "scaling-reinforcement-learning-lessons-from-trillion-token-deployments-at-fortune-500s",
    "title": "Scaling Reinforcement Learning: Lessons from Trillion-Token Deployments at Fortune 500s",
    "url": "https://ai.engineer/talks/scaling-reinforcement-learning-lessons-from-trillion-token-deployments-at-fortune-500s",
    "videoId": "X6NShR2ccOg",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1115000,
    "speakers": [
      {
        "slug": "alessandro-cappelli",
        "name": "Alessandro Cappelli",
        "organization": "Adaptive ML"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Adaptive ML co-founder Alessandro Cappelli argues that reinforcement learning closes the gap between enterprise AI demonstrations and production by continuously incorporating operational feedback. He contrasts RL with supervised fine-tuning, explains how smaller specialized models improve inference cost and latency at enterprises including AT&T, and describes training agents in existing environments while generating synthetic trajectories and bootstrapping datasets through rejection sampling. The presentation concludes with an audience question about human feedback."
  },
  {
    "slug": "scaling-the-next-paradigm-of-heterogeneous-intelligence",
    "title": "Scaling the Next Paradigm of Heterogeneous Intelligence",
    "url": "https://ai.engineer/talks/scaling-the-next-paradigm-of-heterogeneous-intelligence",
    "videoId": "WRBNDpUhsJQ",
    "event": "AI Engineer Europe 2026",
    "durationMs": 913000,
    "speakers": [
      {
        "slug": "adrian-bertagnoli",
        "name": "Adrian Bertagnoli",
        "organization": "Callosum"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Callosum founding engineer Adrian Bertagnoli argues that AI systems should jointly optimize model architectures, agent workflows, and heterogeneous hardware rather than scale one model across identical chips. He describes recursive language models that access long contexts through a Python REPL and reports an OOLONG benchmark configuration on Cerebras that is seven times cheaper and five times faster. A visual web-navigation case study combines Qwen3-VL-8B-Instruct with Kimi K2.5 or GPT models, reporting accuracy, latency, and cost improvements through task-specific model routing; the presentation concludes with an audience question about how routing decisions are made."
  },
  {
    "slug": "scaling-to-long-horizons",
    "title": "Scaling to Long Horizons",
    "url": "https://ai.engineer/talks/scaling-to-long-horizons",
    "videoId": "2bvtay8wGYI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1087000,
    "speakers": [
      {
        "slug": "ross-taylor",
        "name": "Ross Taylor",
        "organization": "General Reasoning"
      },
      {
        "slug": "chengxi-taylor",
        "name": "Chengxi Taylor",
        "organization": "General Reasoning"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "General Reasoning co-founders Ross Taylor and Chengxi Taylor discuss how reinforcement learning, training environments, compute, and optimization objectives enable agents to tackle long-horizon tasks. Ross traces lessons from Papers With Code, Meta, Galactica, Llama post-training, and InstructGPT, arguing that RLHF turns capable base models into useful products. Chengxi outlines the remaining challenges, including realistic simulation, inference-time and training trade-offs, GPU utilization, and value-model bias."
  },
  {
    "slug": "scaling-up-continual-learning",
    "title": "Scaling up Continual Learning",
    "url": "https://ai.engineer/talks/scaling-up-continual-learning",
    "videoId": "zL1kLftVTlo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1383000,
    "speakers": [
      {
        "slug": "ronak-malde",
        "name": "Ronak Malde",
        "organization": "Trajectory"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Trajectory co-founder Ronak Malde argues that continual learning should convert real-world model interactions into training signals instead of relying primarily on increasingly costly, disconnected benchmarks. He compares SFT, DPO, RLHF, and GRPO, then describes on-policy self-distillation as a richer, token-level teacher-student alternative to scalar reinforcement-learning rewards. The talk discusses distribution shifts, reward hacking, practical OpenClaw-RL integration, and deployment questions about continual-training latency."
  },
  {
    "slug": "second-order-effects",
    "title": "Second Order Effects",
    "url": "https://ai.engineer/talks/second-order-effects",
    "videoId": "apZXpgU_klw",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1306000,
    "speakers": [
      {
        "slug": "cheng-lou",
        "name": "Cheng Lou",
        "organization": "Midjourney"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Midjourney's Cheng Lou explores how AI's most consequential effects can emerge indirectly from changes in computing, learning, creativity, and communication. Using GPU shader art, chess and Go, Conway's Game of Life, speculative musical instruments, Anthropic's sparse-autoencoder research, augmented reality, and Smalltalk-style message passing, he argues that AI can expand human skills and information bandwidth while reshaping user interfaces beyond simple automation."
  },
  {
    "slug": "securing-agents-with-open-standards",
    "title": "Securing Agents with Open Standards",
    "url": "https://ai.engineer/talks/securing-agents-with-open-standards",
    "videoId": "FZoMSupg37E",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1121000,
    "speakers": [
      {
        "slug": "bobby-tiernay",
        "name": "Bobby Tiernay",
        "organization": "Auth0"
      },
      {
        "slug": "kam-sween",
        "name": "Kam Sween",
        "organization": "Auth0"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Auth0 engineers Bobby Tiernay and Kam Sween explain how production AI agents can avoid excessive permissions, exposed secrets, and untraceable actions by tying authorization to individual users. They describe short-lived OAuth tokens, token exchange, fine-grained retrieval-layer authorization, and CIBA-based asynchronous approvals, then demonstrate standards-based agent authorization and Auth0 Token Vault."
  },
  {
    "slug": "security-firewall-for-agents",
    "title": "Security Firewall for Agents",
    "url": "https://ai.engineer/talks/security-firewall-for-agents",
    "videoId": "MkRYPFIMCSA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1146000,
    "speakers": [
      {
        "slug": "ryan-dahl",
        "name": "Ryan Dahl",
        "organization": "Deno"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Deno CEO Ryan Dahl explains how production incident-response agents receive broad access to Postgres, Kubernetes, ClickHouse, AWS, GitHub and Slack, creating destructive-action risks when support inputs prompt-inject otherwise aligned models. Because subprocesses can tunnel non-HTTP Postgres traffic through an EKS endpoint and evade MCP-level controls, Deno built Claw Patrol, an MIT-licensed external firewall that inspects protocol traffic, applies HCL policies, and supports authentication mechanisms including OAuth and AWS SigV4. Dahl argues that increasingly capable models still require independent security backstops; the talk concludes with audience questions."
  },
  {
    "slug": "see-hear-speak-draw",
    "title": "See, Hear, Speak, Draw",
    "url": "https://ai.engineer/talks/see-hear-speak-draw",
    "videoId": "bNZV9s3_u44",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1123000,
    "speakers": [
      {
        "slug": "logan-kilpatrick",
        "name": "Logan Kilpatrick",
        "organization": "OpenAI"
      },
      {
        "slug": "simon-fishman",
        "name": "Simón Fishman",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "OpenAI presenters Logan Kilpatrick and Simón Fishman explain how developers can assemble multimodal applications from specialized models, using text to connect vision, speech transcription, and image generation. They demonstrate a GPT-4V-to-DALL·E 3 workflow that describes a real image and generates a synthetic reinterpretation, then introduce a video-oriented demonstration and discuss more unified multimodal models."
  },
  {
    "slug": "self-driving-products-product-signals-to-pull-requests",
    "title": "Self Driving Products: Product Signals to Pull Requests",
    "url": "https://ai.engineer/talks/self-driving-products-product-signals-to-pull-requests",
    "videoId": "zMiSRliEzv4",
    "event": "AI Engineer Europe 2026",
    "durationMs": 939000,
    "speakers": [
      {
        "slug": "joshua-snyder",
        "name": "Joshua Snyder",
        "organization": "PostHog"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Joshua Snyder of PostHog describes a pipeline that turns product analytics, errors, session replays, and customer messages into automatically generated GitHub pull requests. Signals are ingested, screened for unsafe instructions, normalized, semantically grouped through LLM-generated queries and embeddings, and investigated by a research agent before actionable fixes are implemented and iterated until ready. The agent runs with the Claude Agent SDK in Modal sandboxes, retrieves additional context through an MCP server, and must be evaluated against representative production data."
  },
  {
    "slug": "self-evolving-code-with-ai-enhancing-quality-and-security-in-ci",
    "title": "Self-Evolving Code with AI: Enhancing Quality and Security in CI",
    "url": "https://ai.engineer/talks/self-evolving-code-with-ai-enhancing-quality-and-security-in-ci",
    "videoId": "prttz9_1XDc",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1107000,
    "speakers": [
      {
        "slug": "gunjan-patel",
        "name": "Gunjan Patel",
        "organization": "Palo Alto Networks"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Gunjan Patel of Palo Alto Networks presents a personal, noncommercial approach to self-evolving code that moves time-consuming AI-assisted development work into CI. Contrasting responsive coding copilots with an asynchronous ghostpilot, he describes improving comments and variable names, generating context-aware unit tests and corner cases, incorporating historical Jira bugs, and prioritizing potential issues through simulated AI roles. Proposed fixes remain subject to human review, with GitLab CI and GitHub Actions identified as implementation targets."
  },
  {
    "slug": "self-training-agents-hermes-agent-hf-traces-skills-mcp-finetuning-merve-noyan-hugging-face",
    "title": "Self-Training Agents: Hermes Agent, HF Traces, Skills, MCP & Finetuning — Merve Noyan, Hugging Face",
    "url": "https://ai.engineer/talks/self-training-agents-hermes-agent-hf-traces-skills-mcp-finetuning-merve-noyan-hugging-face",
    "videoId": "OV56RddyFuU",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1151000,
    "speakers": [
      {
        "slug": "merve-noyan",
        "name": "Merve Noyan",
        "organization": "Hugging Face"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Hugging Face's Merve Noyan surveys an open agent ecosystem spanning open-weight and multimodal models, Hugging Face Hub benchmark discovery, inference-provider routing, and local serving tools. She describes Hermes Agent's memory and messaging integrations, explains how MCP and agent skills connect models to practical workflows, demonstrates asking Claude Code to fine-tune a vision-language model, and introduces Hugging Face Storage Buckets for training and agent infrastructure."
  },
  {
    "slug": "semantic-blindness-500-000-sensors-confused-an-llm-raahul-singh-vanc-levstik-phaidra",
    "title": "Semantic Blindness: 500,000 Sensors Confused an LLM - Raahul Singh & Vanč Levstik, Phaidra",
    "url": "https://ai.engineer/talks/semantic-blindness-500-000-sensors-confused-an-llm-raahul-singh-vanc-levstik-phaidra",
    "videoId": "EUsPvBeIx70",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 985000,
    "speakers": [
      {
        "slug": "raahul-singh",
        "name": "Raahul Singh",
        "organization": "Phaidra"
      },
      {
        "slug": "vanc-levstik",
        "name": "Vanč Levstik",
        "organization": "Phaidra"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Phaidra engineers Raahul Singh and Vanč Levstik explain why general-purpose LLMs struggle with large industrial equipment inventories: finite context windows, nearly identical sensor names, unreliable vector retrieval, repetition penalties, and hallucinations from sharded requests. They describe using pre-indexed infrastructure hierarchies and bounded orchestration to preserve operational reliability, reporting approximately 9,000 tokens per query across systems ranging from 64 to 460,000 GPUs. Their broader argument is that production AI-native software should progressively replace unreliable prompted behavior with deterministic implementations where appropriate."
  },
  {
    "slug": "serving-2-million-models-without-melting-scaling-the-hugging-face-hub",
    "title": "Serving 2 Million Models Without Melting: Scaling the Hugging Face Hub",
    "url": "https://ai.engineer/talks/serving-2-million-models-without-melting-scaling-the-hugging-face-hub",
    "videoId": "lyL5QhgIOxc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1299000,
    "speakers": [
      {
        "slug": "arek-borucki",
        "name": "Arek Borucki",
        "organization": "Hugging Face"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Hugging Face engineer Arek Borucki explains how the Hub supports a rapidly growing catalog of approximately three million models and more than 14 million users. He describes prioritizing P99 search latency, separating MongoDB Atlas metadata from model artifacts in AWS S3, using denormalized read collections and Apache Lucene-powered Atlas Search for tokenized autocomplete and relevance ranking, and scaling Kubernetes workloads with Horizontal Pod Autoscaler."
  },
  {
    "slug": "serving-voice-ai-at-1-hr-open-source-loras-latency-load-balancing",
    "title": "Serving Voice AI at $1/hr: Open-source, LoRAs, Latency, Load Balancing",
    "url": "https://ai.engineer/talks/serving-voice-ai-at-1-hr-open-source-loras-latency-load-balancing",
    "videoId": "rD23-VZZHOo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 969000,
    "speakers": [
      {
        "slug": "neil-dwyer",
        "name": "Neil Dwyer",
        "organization": "Gabber"
      },
      {
        "slug": "jack-dwyer",
        "name": "Jack Dwyer",
        "organization": "Gabber"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Gabber CTO Neil Dwyer explains how the company deployed the open-source Orpheus text-to-speech model for affordable, real-time consumer voice AI. He discusses voice demonstrations, latency metrics and leading silence, GPU deployment, vLLM-based batched inference with LoRA adapters, load balancing, and the surrounding LiveKit, WebRTC, and SNAC ecosystem."
  },
  {
    "slug": "serving-voice-ai-at-scale-arjun-desai-cartesia-rohit-talluri-aws",
    "title": "Serving Voice AI at Scale — Arjun Desai (Cartesia) & Rohit Talluri (AWS)",
    "url": "https://ai.engineer/talks/serving-voice-ai-at-scale-arjun-desai-cartesia-rohit-talluri-aws",
    "videoId": "knH3fmGAteQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1025000,
    "speakers": [
      {
        "slug": "arjun-desai",
        "name": "Arjun Desai",
        "organization": "Cartesia"
      },
      {
        "slug": "rohit-talluri",
        "name": "Rohit Talluri",
        "organization": "Amazon Web Services (AWS)"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "AWS's Rohit Talluri interviews Cartesia co-founder Arjun Desai about serving enterprise voice AI at scale. Desai explains real-time multimodal foundation models, state-space-model inference, Sonic 2, latency tradeoffs across speech-to-text and language-model pipelines, voice customization and cloning, and edge deployment. Audience questions extend the discussion to video-model research and local models."
  },
  {
    "slug": "shift-left-how-to-become-an-ai-engineer-from-a-full-stack-background",
    "title": "Shift Left: How to Become an AI Engineer from a Full-Stack Background",
    "url": "https://ai.engineer/talks/shift-left-how-to-become-an-ai-engineer-from-a-full-stack-background",
    "videoId": "a0NwFm5NxGo",
    "event": "AI Engineer Summit 2023",
    "durationMs": 619000,
    "speakers": [
      {
        "slug": "reid-mayo",
        "name": "Reid Mayo",
        "organization": "re:ma AI"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Reid Mayo presents a practical syllabus for experienced full-stack developers entering AI engineering without prior machine-learning experience. He recommends studying large language models and transformers, using ChatGPT as a Socratic tutor, practicing prompt engineering, learning OpenAI APIs and examples, and building modular applications with LangChain and retrieval-augmented generation. The curriculum also introduces evaluation techniques, fine-tuning, and model training, with resources from Cohere, Prompt Engineering Guide, Learn Prompting, and the OpenAI Cookbook."
  },
  {
    "slug": "ship-agents-that-ship-a-hands-on-workshop-for-swe-agent-builders",
    "title": "Ship Agents that Ship: A Hands-On Workshop for SWE Agent Builders",
    "url": "https://ai.engineer/talks/ship-agents-that-ship-a-hands-on-workshop-for-swe-agent-builders",
    "videoId": "Fzb1a24hF-o",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 4860000,
    "speakers": [
      {
        "slug": "kyle-penfound",
        "name": "Kyle Penfound",
        "organization": "DAgger"
      },
      {
        "slug": "jeremy-adams-casanas",
        "name": "Jeremy Adams - Casañas",
        "organization": "DAgger"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Dagger presenters Kyle Penfound and Jeremy Adams demonstrate how to build a production-minded software engineering agent, starting from a Hello Dagger application template and giving the agent narrowly scoped tools for inspecting a workspace, editing files, and running tests. They discuss balancing autonomy against reliability, demonstrate Claude-assisted code changes and a GitHub issue-label-triggered workflow, and examine TypeScript implementations, nested container execution, and more advanced example repositories. Audience questions and intermittent workshop connectivity problems are also captured."
  },
  {
    "slug": "ship-it-building-production-ready-agents",
    "title": "Ship it! Building Production-Ready Agents",
    "url": "https://ai.engineer/talks/ship-it-building-production-ready-agents",
    "videoId": "HT4l0DeP69I",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1177000,
    "speakers": [
      {
        "slug": "mike-chambers",
        "name": "Mike Chambers",
        "organization": "Amazon Web Services"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "AWS developer advocate Mike Chambers demonstrates how to take a simple Python dice-rolling AI agent into production. He explains agent prompts, tool examples, conversational history, AWS-hosted model choices and Amazon Bedrock action groups, then discusses infrastructure-as-code deployment and Amazon Q Developer-assisted implementation."
  },
  {
    "slug": "ship-production-software-in-minutes-not-months",
    "title": "Ship Production Software in Minutes, Not Months",
    "url": "https://ai.engineer/talks/ship-production-software-in-minutes-not-months",
    "videoId": "iheWKg2Tkrk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 966000,
    "speakers": [
      {
        "slug": "eno-reyes",
        "name": "Eno Reyes",
        "organization": "Factory"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Factory's Eno Reyes argues that enterprise software engineering should move beyond incremental AI coding assistance and vibe coding toward agents operating across the development lifecycle. He demonstrates how Droids inspect repositories, clarify requirements, write and validate code, and produce CI-ready pull requests; emphasizes organizational context over model capability alone; and describes using MCP-connected tools, meeting transcripts, and Knowledge Droid to create plans, PRDs, roadmaps, incident-response workflows, and runbooks."
  },
  {
    "slug": "ship-real-agents-hands-on-evals-for-agentic-applications",
    "title": "Ship Real Agents: Hands-On Evals for Agentic Applications",
    "url": "https://ai.engineer/talks/ship-real-agents-hands-on-evals-for-agentic-applications",
    "videoId": "Xfl50508LZM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 7458000,
    "speakers": [
      {
        "slug": "laurie-voss",
        "name": "Laurie Voss",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Arize AI’s Laurie Voss leads a hands-on workshop on evaluating a Claude-powered financial-analysis agent. He explains why agent evaluation begins with [REDACTED:username] tracing and failure analysis, then develops deterministic code checks, faithfulness evaluations, and custom LLM-as-a-judge rubrics. The workshop emphasizes choosing appropriate evaluators, distinguishing capability from regression tests, preventing agents from gaming evaluations, and improving prompts through measured iteration."
  },
  {
    "slug": "shipping-ai-that-works-an-evaluation-framework-for-pms",
    "title": "Shipping AI That Works: An Evaluation Framework for PMs",
    "url": "https://ai.engineer/talks/shipping-ai-that-works-an-evaluation-framework-for-pms",
    "videoId": "2HNSG990Ew8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 5176000,
    "speakers": [
      {
        "slug": "aman-khan",
        "name": "Aman Khan",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Arize AI product manager Aman Khan leads an interactive workshop on building repeatable evaluation workflows for AI products. Drawing on experience at Cruise and Spotify, he demonstrates prompt editing, experiment comparisons, agent tool-call evaluation, and response-quality assessment. Audience questions explore LLM-as-a-judge variance, temperature, BERT and ALBERT, human-labeled data, evaluation-team hiring, and shared ownership of prompts and evaluations across product and engineering."
  },
  {
    "slug": "shipping-ai-to-a-million-patients-without-an-a-b-test",
    "title": "Shipping AI to a Million Patients Without an A/B Test",
    "url": "https://ai.engineer/talks/shipping-ai-to-a-million-patients-without-an-a-b-test",
    "videoId": "McknwOzbmyg",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1155000,
    "speakers": [
      {
        "slug": "jared-joselowitz",
        "name": "Jared Joselowitz",
        "organization": "Ufonia"
      }
    ],
    "topics": [
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Ufonia research engineer Jared Joselowitz explains why conventional A/B testing, incremental rollout, and rollback cannot adequately protect patients interacting with Dora, a regulated clinical voice agent. He presents MATRIX, a hazard-driven simulation framework in which the LLM-based PathBot generates realistic patient conversations, and describes validating simulated-patient realism and optimizing prompts with rewards weighted toward detecting clinical red flags before deployment."
  },
  {
    "slug": "shipping-an-enterprise-voice-ai-agent-in-100-days",
    "title": "Shipping an Enterprise Voice AI Agent in 100 Days",
    "url": "https://ai.engineer/talks/shipping-an-enterprise-voice-ai-agent-in-100-days",
    "videoId": "HOYLZ7IVgJo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1030000,
    "speakers": [
      {
        "slug": "peter-bar",
        "name": "Peter Bar",
        "organization": "Intercom"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Intercom product leader Peter Bar explains how Fin Voice was developed in approximately 100 days as an enterprise phone-support agent that answers customer questions and escalates to humans when necessary. The talk covers selecting high-value support use cases, scoping an MVP around knowledge-base answers, authentication and routing, comparing chained speech-to-text/LLM/text-to-speech architectures with direct voice-to-voice models, and building call observability and integrations that fit existing support-team workflows."
  },
  {
    "slug": "shipping-complex-ai-applications-braintrust-trainline",
    "title": "Shipping complex AI applications | Braintrust & Trainline",
    "url": "https://ai.engineer/talks/shipping-complex-ai-applications-braintrust-trainline",
    "videoId": "ZdheJTfLu-s",
    "event": "AI Engineer Europe 2026",
    "durationMs": 5914000,
    "speakers": [
      {
        "slug": "giran-moodley",
        "name": "Giran Moodley",
        "organization": "Braintrust"
      },
      {
        "slug": "mayan-soni",
        "name": "Mayan Soni"
      },
      {
        "slug": "oussama-hafferssas",
        "name": "Oussama Hafferssas",
        "organization": "Trainline"
      },
      {
        "slug": "mayank-soni",
        "name": "Mayank Soni",
        "organization": "Trainline"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Braintrust and Trainline present a hands-on workshop on shipping reliable, production-grade AI applications. Giran Moodley, Oussama Hafferssas, and Mayank Soni discuss Trainline’s agentic systems, decomposing LLM workflows into tool-enabled specialist stages, instrumenting applications with the Braintrust SDK, tracing calls and diagnosing failures, and continuously evaluating production logs through automated online scoring."
  },
  {
    "slug": "shipping-products-when-you-don-t-know-what-they-can-do",
    "title": "Shipping Products When You Don’t Know What they Can Do",
    "url": "https://ai.engineer/talks/shipping-products-when-you-don-t-know-what-they-can-do",
    "videoId": "PthmdT92qNg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1174000,
    "speakers": [
      {
        "slug": "ben-stein",
        "name": "Ben Stein",
        "organization": "Teammates"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Teammates founder Ben Stein explains why product management for autonomous AI coworkers must adapt to unpredictable LLM capabilities and emergent behavior. Using a virtual colleague operating across email, Slack, and document comments, he advocates specifying broad affordances instead of rigid features, treating AI evaluations and LLM-based assessments as living product specifications, and developing clearer criteria for testing, triaging, and classifying unexpected agent behavior."
  },
  {
    "slug": "shipping-something-to-someone-always-wins",
    "title": "Shipping something to someone always wins",
    "url": "https://ai.engineer/talks/shipping-something-to-someone-always-wins",
    "videoId": "mHzJhXppwUA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 977000,
    "speakers": [
      {
        "slug": "kenneth-auchenberg",
        "name": "Kenneth Auchenberg",
        "organization": "AlleyCorp"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Kenneth Auchenberg applies lessons from building Visual Studio Code and Stripe’s developer platform to AI-era product management: ship a continuously useful product to real users, tighten feedback loops before irreversible decisions, validate positioning and API designs early, and use AI-assisted prototypes to accelerate iteration. He emphasizes direct customer observation, monitoring integration friction, and taking particular care with APIs and data structures because they are harder to change than user interfaces."
  },
  {
    "slug": "simulation-maxxing-how-nubank-ships-agents-20-faster-with-simulations",
    "title": "Simulation-Maxxing: How Nubank ships agents 20× faster with simulations",
    "url": "https://ai.engineer/talks/simulation-maxxing-how-nubank-ships-agents-20-faster-with-simulations",
    "videoId": "KMR_RBoCa4M",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 989000,
    "speakers": [
      {
        "slug": "shreya-rajpal",
        "name": "Shreya Rajpal",
        "organization": "Snowglobe"
      },
      {
        "slug": "aman-gupta",
        "name": "Aman Gupta",
        "organization": "Nubank"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Shreya Rajpal of Snowglobe and Aman Gupta of Nubank explain how generating synthetic agent-evaluation data through simulations accelerates customer-support agent deployment by a reported 20×. They describe the bottlenecks of hand-curated evaluations and production A/B tests, demonstrate the Snowglobe SDK and simulated customer personas, and discuss human-aligned evaluation metrics, automated prompt optimization, improved customer satisfaction, reduced production risk, and stronger self-service rates."
  },
  {
    "slug": "skill-issue-how-we-used-ai-to-make-agents-actually-good-at-supabase",
    "title": "Skill Issue: How We Used AI to Make Agents Actually Good at Supabase",
    "url": "https://ai.engineer/talks/skill-issue-how-we-used-ai-to-make-agents-actually-good-at-supabase",
    "videoId": "GmAQKINjv1E",
    "event": "AI Engineer Europe 2026",
    "durationMs": 4721000,
    "speakers": [
      {
        "slug": "pedro-rodrigues",
        "name": "Pedro Rodrigues",
        "organization": "Supabase"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Supabase AI tooling engineer Pedro Rodrigues delivers a workshop retitled “Level Up Your Skills,” demonstrating how to build and refine Agent Skills that improve agent behavior on real Supabase workflows. He introduces agent experience, SKILL.md and supporting scripts, a demonstration application and GitHub repository, Claude Code, and Vercel’s skills package. The workshop emphasizes designing evaluation scenarios, inspecting outcomes, and using LLM-as-a-judge techniques to measure whether skill changes actually improve performance."
  },
  {
    "slug": "skill-issue-lessons-from-skilling-up-coding-agents-to-use-langfuse",
    "title": "Skill issue: Lessons from skilling up coding agents to use Langfuse",
    "url": "https://ai.engineer/talks/skill-issue-lessons-from-skilling-up-coding-agents-to-use-langfuse",
    "videoId": "vNCY9kXXyDQ",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1449000,
    "speakers": [
      {
        "slug": "marc-klingen",
        "name": "Marc Klingen",
        "organization": "Langfuse, part of ClickHouse"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Langfuse co-founder Marc Klingen explains how agent skills help coding agents instrument applications reliably despite stale model knowledge. He examines Claude Code tracing failures, progressive documentation references, agent-accessible APIs and CLIs, data-region selection, hallucinated CLI parameters, documentation-search telemetry, evaluation design, LLM-as-a-Judge workflows, and challenges distributing skills. The presentation ends with audience questions."
  },
  {
    "slug": "skills-are-new-features-building-skill-centric-harness-yogendra-miraje-factset",
    "title": "Skills are new features: Building Skill-Centric Harness — Yogendra Miraje, FactSet",
    "url": "https://ai.engineer/talks/skills-are-new-features-building-skill-centric-harness-yogendra-miraje-factset",
    "videoId": "7jjudsEhBtM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1044000,
    "speakers": [
      {
        "slug": "yogendra-miraje",
        "name": "Yogendra Miraje",
        "organization": "FactSet"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "FactSet's Yogendra Miraje argues that reusable agent skills are the new product features and that engineers should build the harnesses that execute them. He describes replacing proprietary blueprints with Anthropic's Agent Skills standard, implementing agentic loops and selective skill loading, routing financial-research and HTML/PDF report-generation tasks through skill descriptions, and governing enterprise skill registries with human-reviewed admission controls."
  },
  {
    "slug": "skills-are-the-new-sdks",
    "title": "Skills are the New SDKs",
    "url": "https://ai.engineer/talks/skills-are-the-new-sdks",
    "videoId": "LC3-P7v3yoI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1600000,
    "speakers": [
      {
        "slug": "elvin-aghammadzada",
        "name": "Elvin Aghammadzada",
        "organization": "DataRobot"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Elvin Aghammadzada of DataRobot argues that enterprise platforms need versioned, task-specific agent skills—not merely APIs, SDKs, documentation, or large MCP tool catalogs—to make their workflows teachable to coding agents. The presentation examines context-window degradation, context engineering, metadata-based skill discovery, enterprise governance, and how skills complement or sometimes overlap with MCP. It concludes that skills should be treated as production software requiring evaluation, testing, and versioning; a second unidentified participant briefly contributes to the discussion."
  },
  {
    "slug": "skills-at-scale",
    "title": "Skills at Scale",
    "url": "https://ai.engineer/talks/skills-at-scale",
    "videoId": "pFsfax19yOM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 4863000,
    "speakers": [
      {
        "slug": "nick-nisi",
        "name": "Nick Nisi",
        "organization": "WorkOS"
      },
      {
        "slug": "zack-proser",
        "name": "Zack Proser",
        "organization": "WorkOS"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "WorkOS engineers Nick Nisi and Zack Proser lead an interactive workshop on designing portable, composable AI-agent skills. They build a Repo Roast example, explain selective context loading and reusable workflow scripts, demonstrate Claude-to-Codex review and public WorkOS migration references, and show how the WorkOS CLI automates AuthKit integration while fielding audience questions."
  },
  {
    "slug": "small-ai-teams-with-huge-impact-vik-paruchuri-datalab",
    "title": "Small AI Teams with Huge Impact — Vik Paruchuri, Datalab",
    "url": "https://ai.engineer/talks/small-ai-teams-with-huge-impact-vik-paruchuri-datalab",
    "videoId": "K-iYKDMFKhE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1056000,
    "speakers": [
      {
        "slug": "vikas-paruchuri",
        "name": "Vikas Paruchuri",
        "organization": "Datalab"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Datalab CEO Vikas Paruchuri explains how a small, generalist AI team built the Marker and Surya document-intelligence projects, grew revenue, and trained OCR models without extensive organizational specialization. Drawing on Jeremy Howard's approach at Answer.AI, he advocates end-to-end ownership, AI-assisted tooling, fewer handoffs, reusable components, and a simple server-rendered technology stack. The talk concludes with audience questions about organizational scaling and hiring."
  },
  {
    "slug": "small-bets-big-impact-building-genbi-at-a-fortune-100",
    "title": "Small Bets, Big Impact: Building GenBI at a Fortune 100",
    "url": "https://ai.engineer/talks/small-bets-big-impact-building-genbi-at-a-fortune-100",
    "videoId": "LU9KgcZDRfY",
    "event": "AI Engineer Code 2025",
    "durationMs": 1370000,
    "speakers": [
      {
        "slug": "asaf-bord",
        "name": "Asaf Bord",
        "organization": "Northwestern Mutual"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Asaf Bord explains how Northwestern Mutual is developing GenBI, an AI-powered business-intelligence agent, through incremental projects that earn funding and trust inside a risk-averse enterprise. The approach starts with real enterprise data, existing verified reports, dashboards, metadata, and BI practitioners; rolls out first to BI experts and then business managers; and adds governance, orchestration, and contextual interfaces while managing production complexity and competition from products such as Databricks Genie."
  },
  {
    "slug": "software-development-agents-what-works-and-what-doesn-t",
    "title": "Software Development Agents: What Works and What Doesn't",
    "url": "https://ai.engineer/talks/software-development-agents-what-works-and-what-doesn-t",
    "videoId": "o_hhkJtlbSs",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1006000,
    "speakers": [
      {
        "slug": "robert-brennan",
        "name": "Robert Brennan",
        "organization": "All Hands AI / OpenHands"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "OpenHands cocreator Robert Brennan explains when autonomous software-development agents outperform autocomplete, how humans retain responsibility for product judgment and broader engineering decisions, and how agents combine LLM-driven action loops with editors, terminals, browsers, and sandboxes. He discusses efficient file edits, browser representations, safe execution, and practical starting points such as addressing pull-request feedback and greenfield work."
  },
  {
    "slug": "software-engineering-ai",
    "title": "Software Engineering + AI = ?",
    "url": "https://ai.engineer/talks/software-engineering-ai",
    "videoId": "CS5Cmz5FssI",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1602000,
    "speakers": [
      {
        "slug": "gergely-orosz",
        "name": "Gergely Orosz",
        "organization": "The Pragmatic Engineer"
      },
      {
        "slug": "swyx",
        "name": "swyx",
        "organization": "AI Engineer"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Gergely Orosz joins an AI Engineer Europe fireside conversation about how AI coding tools are changing software engineering. They examine token-maxing leaderboards, token-spend incentives, performance evaluation, unreliable productivity metrics, skepticism toward AI adoption, and leadership pressure inside large companies. The discussion also covers Cursor, Shopify’s early engagement with GitHub Copilot, and the growth of The Pragmatic Engineer into a paid newsletter."
  },
  {
    "slug": "software-engineering-is-becoming-plan-and-review",
    "title": "Software Engineering Is Becoming Plan and Review",
    "url": "https://ai.engineer/talks/software-engineering-is-becoming-plan-and-review",
    "videoId": "W76woOYHlvY",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1223000,
    "speakers": [
      {
        "slug": "louis-knight-webb",
        "name": "Louis Knight-Webb",
        "organization": "Vibe Kanban"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Vibe Kanban founder Louis Knight-Webb argues that increasingly capable coding agents are shifting software engineering away from manual implementation toward planning and reviewing AI-generated work. He contrasts plan-heavy and review-heavy workflows, describes why front-end features remain difficult to specify, and highlights longer agent runtimes, browser-based QA with Playwright MCP, and the behavioral consequences of crossing a five-minute execution threshold. He also announces that Vibe Kanban is shutting down and reflects on startup strategy and enterprise sales."
  },
  {
    "slug": "sovereign-escape-velocity-ownership-with-open-models-gus-martins-and-ian-ballantyne-google-deepm",
    "title": "Sovereign Escape Velocity: Ownership with Open Models — Gus Martins and Ian Ballantyne, Google DeepMind",
    "url": "https://ai.engineer/talks/sovereign-escape-velocity-ownership-with-open-models-gus-martins-and-ian-ballantyne-google-deepm",
    "videoId": "SS-A8sE7hkw",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1252000,
    "speakers": [
      {
        "slug": "gus-martins",
        "name": "Gus Martins",
        "organization": "Google DeepMind"
      },
      {
        "slug": "ian-ballantyne",
        "name": "Ian Ballantyne",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Google DeepMind's Gus Martins and Ian Ballantyne explain how Gemma 4 open models enable greater control over data, infrastructure, deployment, and model customization than hosted proprietary alternatives. They highlight efficient phone and single-GPU deployment, benchmark performance relative to model size, language-specific adaptations, and the economics of running increasingly token-intensive coding and agentic workloads on local hardware."
  },
  {
    "slug": "spec-driven-development-agentic-coding-at-faang-scale-and-quality-al-harris-amazon-kiro",
    "title": "Spec-Driven Development: Agentic Coding at FAANG Scale and Quality — Al Harris, Amazon Kiro",
    "url": "https://ai.engineer/talks/spec-driven-development-agentic-coding-at-faang-scale-and-quality-al-harris-amazon-kiro",
    "videoId": "HY_JyxAZsiE",
    "event": "AI Engineer Code 2025",
    "durationMs": 3830000,
    "speakers": [
      {
        "slug": "al-harris",
        "name": "Al Harris",
        "organization": "Amazon"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Amazon principal engineer Al Harris demonstrates how Kiro uses spec-driven development to turn prompts into structured requirements, acceptance criteria, design artifacts, and implementation tasks. He contrasts this workflow with vibe coding, discusses MCP integrations, steering documents, context management, large codebases, multi-root workspaces, and future sub-agent parallelism, and demonstrates deploying a sample joke-generating agent to Amazon Bedrock AgentCore while answering audience questions."
  },
  {
    "slug": "spec-driven-testing-for-agents-with-a-brain-the-size-of-a-planet-steven-willmott-safe-intelligen",
    "title": "Spec-Driven Testing for Agents With A Brain the Size of A Planet — Steven Willmott, Safe Intelligence",
    "url": "https://ai.engineer/talks/spec-driven-testing-for-agents-with-a-brain-the-size-of-a-planet-steven-willmott-safe-intelligen",
    "videoId": "UQKg0td-Bf4",
    "event": "AI Engineer Europe 2026",
    "durationMs": 783000,
    "speakers": [
      {
        "slug": "steven-willmott",
        "name": "Steven Willmott",
        "organization": "Safe Intelligence"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Safe Intelligence CEO Steven Willmott argues that increasingly capable AI agents require implementation-independent behavioral specifications rather than evaluation datasets alone. Drawing on formal verification and robustness testing, he discusses jailbreak exposure, expanding attack surface, explicit customer-service constraints, domain-specific distinctions, input variation, and reusable specifications that remain useful when agent frameworks or models change."
  },
  {
    "slug": "special-topics-in-kernels-rl-reward-hacking-in-agents",
    "title": "Special topics in Kernels, RL, Reward Hacking in Agents",
    "url": "https://ai.engineer/talks/special-topics-in-kernels-rl-reward-hacking-in-agents",
    "videoId": "uIiA6DquRiE",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 8421000,
    "speakers": [
      {
        "slug": "daniel-han",
        "name": "Daniel Han",
        "organization": "Unsloth"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Unsloth co-founder Daniel Han presents an extended technical workshop on open-model distribution, training-stack improvements, AI capability trends, dynamic low-bit quantization, and cost-performance tradeoffs. He examines benchmark reliability through SWE-bench and mathematical answer verification, discusses cybersecurity and regulation, and contrasts software optimization with specialized hardware and GPU-kernel development. Audience questions appear throughout the recording."
  },
  {
    "slug": "spreadsheets-are-all-you-need-decoding-the-decoder-llm-without-de-code",
    "title": "Spreadsheets-are-all-you-need: Decoding the Decoder LLM without de code",
    "url": "https://ai.engineer/talks/spreadsheets-are-all-you-need-decoding-the-decoder-llm-without-de-code",
    "videoId": "NamKkerrlnQ",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1028000,
    "speakers": [
      {
        "slug": "ishan-anand",
        "name": "Ishan Anand"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Ishan Anand demonstrates a complete GPT-2 Small implementation in Excel to explain tokenization, 768-dimensional embeddings, matrix multiplication, normalization, logits, and next-token prediction without Python or API calls. He then traces intermediate representations across model layers and demonstrates interpretability and activation steering using sparse-autoencoder features, Anthropic's Golden Gate Claude example, and representation engineering."
  },
  {
    "slug": "state-of-data",
    "title": "State of Data",
    "url": "https://ai.engineer/talks/state-of-data",
    "videoId": "ZyIoTOAbRfs",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1102000,
    "speakers": [
      {
        "slug": "sean-cai",
        "name": "Sean Cai",
        "organization": "Independent / State of Data"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Sean Cai examines how AI training-data markets are shifting from static annotations and final-state records toward authentic enterprise workflow and process data. He distinguishes Type 1 from Type 2 data, links application maturity to task verifiability, and argues that benchmark gaming and evaluation-harness differences obscure real model capability. Finance-task examples illustrate the need for robust rubrics, deterministic verification, and durable enterprise data infrastructure."
  },
  {
    "slug": "state-of-startups-and-ai-2025",
    "title": "State of Startups and AI 2025",
    "url": "https://ai.engineer/talks/state-of-startups-and-ai-2025",
    "videoId": "3MZS5gNElZM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1432000,
    "speakers": [
      {
        "slug": "sarah-guo",
        "name": "Sarah Guo",
        "organization": "Conviction"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Conviction investor Sarah Guo surveys AI-native startups and venture investing, contrasting hype with sustained product adoption and revenue. She discusses production coding agents, voice AI and inference economics; defines agents through planning, task ownership, and goal memory; highlights companies spanning models, infrastructure, and applications; and argues that practical copilots, domain expertise, and applied machine learning remain important opportunities."
  },
  {
    "slug": "state-of-the-union-why-local-why-now",
    "title": "State of the Union: Why Local, Why Now",
    "url": "https://ai.engineer/talks/state-of-the-union-why-local-why-now",
    "videoId": "KB41dTlX1Uc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2669000,
    "speakers": [
      {
        "slug": "nader-khalil",
        "name": "Nader Khalil",
        "organization": "NVIDIA"
      },
      {
        "slug": "alex-cheema",
        "name": "Alex Cheema",
        "organization": "EXO Labs"
      },
      {
        "slug": "matthew-berman",
        "name": "Matthew Berman",
        "organization": "Forward Future"
      },
      {
        "slug": "ahmad-osman",
        "name": "Ahmad Osman",
        "organization": "Osmantic"
      },
      {
        "slug": "joseph-nelson",
        "name": "Joseph Nelson",
        "organization": "Roboflow"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "NVIDIA moderator Nader Khalil joins Alex Cheema of EXO Labs, Matthew Berman of Forward Future, Ahmad Osman of Osmantic, and Joseph Nelson of Roboflow to examine why local AI is reaching an adoption inflection point. The panel discusses distributed inference on NVIDIA DGX systems, computer-vision applications, multi-model routing and context management, simpler onboarding, specialized models and fine-tuning, and the remaining technical and open-source-access challenges."
  },
  {
    "slug": "state-space-models-for-realtime-multimodal-intelligence",
    "title": "State Space Models for Realtime Multimodal Intelligence",
    "url": "https://ai.engineer/talks/state-space-models-for-realtime-multimodal-intelligence",
    "videoId": "U9DPRZ0lSIQ",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 866000,
    "speakers": [
      {
        "slug": "karan-goel",
        "name": "Karan Goel",
        "organization": "Cartesia"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Cartesia's Karan Goel contrasts slow, cloud-based batch AI with streaming multimodal systems that must respond instantly on low-power devices. He describes state space models and Mamba as alternatives to transformer architectures that repeatedly retain and inspect extensive context, emphasizing compressed internal memory, long-lived intelligence, and streaming token updates. The talk closes with Cartesia's low-latency voice generation and efforts to run comparable models efficiently on Macs and other devices."
  },
  {
    "slug": "stateful-agents-full-workshop-with-redacted-username-packer-of-letta-and-memgpt",
    "title": "Stateful Agents — Full Workshop with Packer of Letta and MemGPT",
    "url": "https://ai.engineer/talks/stateful-agents-full-workshop-with-redacted-username-packer-of-letta-and-memgpt",
    "videoId": "E0k9Ppq6yXY",
    "event": "AI Engineer Summit 2025",
    "durationMs": 4774000,
    "speakers": [],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "[REDACTED:username] Packer introduces stateful agents through the MemGPT research and the Letta framework, explaining how persistent agent handles, editable memory blocks, archival storage, and retrieval preserve context across interactions. The hands-on workshop uses a Docker-hosted server and Python notebooks, then explores the Agent Development Environment, ReAct-style reasoning, tool calls, observability, and deployment latency."
  },
  {
    "slug": "stateful-environments-for-vertical-agents-josh-purtell-synth-labs",
    "title": "Stateful environments for vertical agents — Josh Purtell, Synth Labs",
    "url": "https://ai.engineer/talks/stateful-environments-for-vertical-agents-josh-purtell-synth-labs",
    "videoId": "5rMc-moNVx0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 411000,
    "speakers": [
      {
        "slug": "josh-purtell",
        "name": "Josh Purtell",
        "organization": "Synth"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Synth founder Josh Purtell presents stateful environments as a practical abstraction for vertical AI agents operating over long horizons. Drawing on reinforcement-learning environments, OpenAI Gym, SWE-bench, and SWE-agent, he explains how agents can interact with simplified representations of external systems such as Excel while retaining useful state. Stateful environments enable rollback and branching for approaches such as Language Agent Tree Search, illustrated with Minecraft, and he points to Synth’s open-source environment implementations."
  },
  {
    "slug": "stop-ai-agent-hallucinations-5-techniques-production-patterns-elizabeth-fuentes-leone-aws",
    "title": "Stop AI Agent Hallucinations: 5 Techniques + Production Patterns - Elizabeth Fuentes Leone, AWS",
    "url": "https://ai.engineer/talks/stop-ai-agent-hallucinations-5-techniques-production-patterns-elizabeth-fuentes-leone-aws",
    "videoId": "vJukHCIv7Ck",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 3319000,
    "speakers": [
      {
        "slug": "elizabeth-fuentes-leone",
        "name": "Elizabeth Fuentes Leone",
        "organization": "AWS"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "AWS developer advocate Elizabeth Fuentes Leone demonstrates code-level techniques for reducing AI-agent hallucinations and token waste using a synthetic travel-booking agent built with Strands Agents. Examples cover embedding-based semantic tool selection, Neo4j-backed GraphRAG, explicit booking and payment guardrails, and Agent Control SDK steering, with integrations for OpenAI, Amazon Bedrock, and Ollama. She cautions that the simplified demonstration tools are not themselves production-ready."
  },
  {
    "slug": "stop-babysitting-your-agents-building-a-context-engine-for-mergeable-code",
    "title": "Stop babysitting your agents: building a context engine for mergeable code",
    "url": "https://ai.engineer/talks/stop-babysitting-your-agents-building-a-context-engine-for-mergeable-code",
    "videoId": "BiG2ssibKGc",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1134000,
    "speakers": [
      {
        "slug": "brandon-waselnuk",
        "name": "Brandon Waselnuk",
        "organization": "Unblocked"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Unblocked developer advocate Brandon Waselnuk explains why coding agents require dynamically assembled organizational context rather than additional tool connections, static instruction files, or naive retrieval alone. He describes context engines that combine social-graph relevance, token-efficient research packets, permission-aware access to workplace conversations, conflict resolution, and freshness controls, then demonstrates an MCP-assisted Zendesk integration workflow."
  },
  {
    "slug": "stop-burning-tokens-why-self-improvement-needs-domain-expertise-first-annabell-schafer-langfuse",
    "title": "Stop Burning Tokens: Why self-improvement needs domain expertise first - Annabell Schäfer, Langfuse",
    "url": "https://ai.engineer/talks/stop-burning-tokens-why-self-improvement-needs-domain-expertise-first-annabell-schafer-langfuse",
    "videoId": "eAXxdtNlK04",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1059000,
    "speakers": [
      {
        "slug": "annabell-schafer",
        "name": "Annabell Schäfer",
        "organization": "Langfuse"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Langfuse presenter Annabell Schäfer explains why self-improving AI agents require domain-specific target functions and expert involvement, rather than indiscriminate optimization loops. She describes a classification workflow pairing GPT-5.4 nano with Claude Code and Claude Opus 4.8 to cluster errors, propose prompt revisions, and accept improvements only when validation performance increases; one illustrated iteration improves accuracy from 68% to 78%. She extends the approach to deterministic checks, knowledge-base grounding, and human-agent collaboration in specialized domains."
  },
  {
    "slug": "stop-evaluating-models-like-it-s-the-50s-alejandro-vidal-mindmakers",
    "title": "Stop Evaluating Models Like It's the 50s - Alejandro Vidal, Mindmakers",
    "url": "https://ai.engineer/talks/stop-evaluating-models-like-it-s-the-50s-alejandro-vidal-mindmakers",
    "videoId": "O3FEoMYvUf8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1415000,
    "speakers": [
      {
        "slug": "alejandro-vidal",
        "name": "Alejandro Vidal",
        "organization": "Mindmakers"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Mindmakers founder Alejandro Vidal argues that counting correct benchmark answers misses differences in question difficulty, uncertainty, and model behavior. Using Epoch AI benchmark data, he demonstrates how item response theory estimates latent ability, distinguishes similarly scoring Claude and Gemini models, identifies flawed answers and informative questions, reduces evaluation workload, assesses response consistency, and uses residual patterns to fingerprint model families."
  },
  {
    "slug": "stop-guessing-build-robust-ai-with-layered-cot",
    "title": "Stop Guessing: Build Robust AI with Layered CoT",
    "url": "https://ai.engineer/talks/stop-guessing-build-robust-ai-with-layered-cot",
    "videoId": "VTJHR7rQ2KI",
    "event": "AI Engineer Summit 2025",
    "durationMs": 616000,
    "speakers": [
      {
        "slug": "manish-sanwal",
        "name": "Manish Sanwal",
        "organization": "News Corp"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Manish Sanwal, who identifies himself as Director of AI at News Corp, introduces Layered Chain-of-Thought for multi-agent AI systems. He explains how specialized agents improve modularity and fault tolerance, why conventional Chain-of-Thought remains vulnerable to prompt sensitivity and cascading unchecked errors, and how verifying each intermediate reasoning step can support more transparent, auditable, reproducible, and self-correcting outputs."
  },
  {
    "slug": "stop-making-models-bigger-make-them-behave-kobie-crawford-snorkel",
    "title": "Stop Making Models Bigger, Make Them Behave — Kobie Crawford, Snorkel",
    "url": "https://ai.engineer/talks/stop-making-models-bigger-make-them-behave-kobie-crawford-snorkel",
    "videoId": "TNwJ1LMiENk",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1256000,
    "speakers": [
      {
        "slug": "kobie-crawford",
        "name": "Kobie Crawford",
        "organization": "Snorkel AI"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "finance",
        "name": "Finance"
      }
    ],
    "summary": "Snorkel AI developer advocate Kobie Crawford explains how a four-billion-parameter model trained with reinforcement learning, GRPO, expert-curated data, and the Berkeley-associated rLLM framework can improve financial-analysis tool use instead of relying on greater model size. He contrasts disciplined SQL schema inspection with a 235-billion-parameter Qwen3 model's problematic financial-query behavior and reports that harder multi-table FinQA reasoning improved from 13.9% to 26.6%."
  },
  {
    "slug": "stop-ordering-ai-takeout-a-cookbook-for-winning-when-you-build-in-house",
    "title": "Stop Ordering AI Takeout: A Cookbook for Winning When You Build In-House",
    "url": "https://ai.engineer/talks/stop-ordering-ai-takeout-a-cookbook-for-winning-when-you-build-in-house",
    "videoId": "DjUIecgpYAo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 645000,
    "speakers": [
      {
        "slug": "jan-siml",
        "name": "Jan Siml"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Jan Siml argues that small internal teams can outperform expensive, generalized AI purchases by building deeply for one business-specific job, attributing AI activity to revenue, and proactively recommending useful actions. He describes a two-developer in-house effort that he says generated several million dollars in annual recurring revenue, contrasts business outcomes with offline evaluation metrics, and discusses production cost tradeoffs between OpenAI o3 and GPT-4.1 mini."
  },
  {
    "slug": "stop-renting-your-cognitive-infrastructure",
    "title": "Stop Renting Your Cognitive Infrastructure",
    "url": "https://ai.engineer/talks/stop-renting-your-cognitive-infrastructure",
    "videoId": "Bck7ABCZRZI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 472000,
    "speakers": [
      {
        "slug": "thiyagarajan-maruthavanan",
        "name": "Thiyagarajan Maruthavanan",
        "organization": "Kalmantic Labs"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Thiyagarajan Maruthavanan argues that metered AI APIs can produce unpredictable inference costs, especially when agent loops waste tokens and context. Drawing on his Ultasono application and a migration from Anthropic to NVIDIA DGX Spark hardware, he contrasts rented inference, cloud token factories, and owned local infrastructure. He recommends renting while validating product-market fit, then owning infrastructure when enterprise cost control and rate-limit autonomy become material, and introduces JustInfer, the open-source JustTokenMax optimization project, and his book on inference economics."
  },
  {
    "slug": "stop-using-rag-as-memory",
    "title": "Stop Using RAG as Memory",
    "url": "https://ai.engineer/talks/stop-using-rag-as-memory",
    "videoId": "T5IMo5ntyhA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 422000,
    "speakers": [
      {
        "slug": "daniel-chalef",
        "name": "Daniel Chalef",
        "organization": "Zep"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Daniel Chalef of Zep argues that vector-based RAG and semantic similarity are insufficient for agent memory because irrelevant facts can overwhelm domain-relevant information. He introduces Graphiti's domain-aware temporal knowledge graphs and demonstrates a finance-coach agent using typed business schemas, Zep's SDK, concurrent searches, and node-type filtering across Python, TypeScript, and Go workflows."
  },
  {
    "slug": "stop-writing-tone-instructions-layer-them",
    "title": "Stop Writing Tone Instructions. Layer Them.",
    "url": "https://ai.engineer/talks/stop-writing-tone-instructions-layer-them",
    "videoId": "ij-AU9dpJjc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1257000,
    "speakers": [
      {
        "slug": "isadora-martin-dye",
        "name": "Isadora Martin-Dye",
        "organization": "Isadora & Co | The Bloom House AI"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Isadora Martin-Dye describes a four-layer architecture for dependable AI brand voice: immutable identity and disclosure rules, situational context, example-anchored voice, and a deterministic post-generation veto. Examples from wedding-venue assistants and a missing-person tool show how the layers prevent false claims of physical presence, unsupported certainty, hallucinated availability, insensitive responses, and cross-tenant identity leaks."
  },
  {
    "slug": "storyteller-building-multi-modal-apps-with-ts-modelfusion-lars-grammel-phd",
    "title": "Storyteller: Building Multi-modal Apps with TS & ModelFusion - Lars Grammel, PhD",
    "url": "https://ai.engineer/talks/storyteller-building-multi-modal-apps-with-ts-modelfusion-lars-grammel-phd",
    "videoId": "KnndhPSVF3E",
    "event": "AI Engineer Summit 2023",
    "durationMs": 450000,
    "speakers": [
      {
        "slug": "lars-grammel",
        "name": "Lars Grammel"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Lars Grammel demonstrates Storyteller, a TypeScript and ModelFusion application that turns a spoken prompt into an illustrated, narrated preschool story. A React client and Fastify server coordinate Whisper transcription, rapid GPT-3.5 outline generation, GPT-4-assisted image prompting, Stable Diffusion XL illustration, voice-embedding retrieval, and synthesized character narration. The talk emphasizes parallel execution, incremental client updates, and model-selection tradeoffs required to keep the experience responsive."
  },
  {
    "slug": "strategies-for-llm-evals-guidellm-lm-eval-harness-openai-evals-workshop-taylor-jordan-smith",
    "title": "Strategies for LLM Evals (GuideLLM, lm-eval-harness, OpenAI Evals Workshop) — Taylor Jordan Smith",
    "url": "https://ai.engineer/talks/strategies-for-llm-evals-guidellm-lm-eval-harness-openai-evals-workshop-taylor-jordan-smith",
    "videoId": "89NuzmKokIk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1948000,
    "speakers": [
      {
        "slug": "taylor-jordan-smith",
        "name": "Taylor Jordan Smith",
        "organization": "Red Hat"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Red Hat AI developer advocate Taylor Jordan Smith presents a hands-on workshop on evaluating production LLM systems beyond leaderboard accuracy. She discusses RAG and agent deployments, inference runtimes including SGLang and vLLM, reliability and risk mitigation, then demonstrates GuideLLM request-rate benchmarking and outlines MMLU-Pro, safety, bias, and custom evaluation exercises. An audience participant asks how prompt and model evaluations can be tied to actual production performance."
  },
  {
    "slug": "structuring-a-modern-ai-team",
    "title": "Structuring a modern AI team",
    "url": "https://ai.engineer/talks/structuring-a-modern-ai-team",
    "videoId": "SbUxRluVRwk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1060000,
    "speakers": [
      {
        "slug": "denys-linkov",
        "name": "Denys Linkov",
        "organization": "Wisedocs"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Wisedocs machine-learning leader Denys Linkov explains how to structure cross-functional AI teams around company context, delivery bottlenecks, domain knowledge, and measurable business value. He argues that adaptable generalists and existing domain-experienced teams often outperform reflexive hiring of specialist AI researchers, illustrates the tradeoff with Ampere's wager, and discusses practical model-training requirements, inner and outer team loops, and continuous learning."
  },
  {
    "slug": "structuring-the-unstructured-advanced-document-parsing-for-ai-workflows",
    "title": "Structuring the Unstructured: Advanced Document Parsing for AI Workflows",
    "url": "https://ai.engineer/talks/structuring-the-unstructured-advanced-document-parsing-for-ai-workflows",
    "videoId": "-x5GEVnkuRw",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1241000,
    "speakers": [
      {
        "slug": "cedric-clyburn",
        "name": "Cedric Clyburn",
        "organization": "Red Hat"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Red Hat’s Cedric Clyburn demonstrates Docling, an open-source, locally runnable document-processing toolkit, for converting PDFs and other unstructured enterprise documents into structured Markdown, JSON, HTML, and Pydantic representations. He explains how OCR, vision models, layout-aware parsing, and table and image extraction preserve context while reducing privacy and cost concerns, then demonstrates applications to retrieval-augmented generation and AI agents."
  },
  {
    "slug": "stuffing-context-is-not-memory-updating-weights-is",
    "title": "Stuffing Context is not Memory, Updating Weights is",
    "url": "https://ai.engineer/talks/stuffing-context-is-not-memory-updating-weights-is",
    "videoId": "Jty4s9-Jb78",
    "event": "AI Engineer Code 2025",
    "durationMs": 3764000,
    "speakers": [
      {
        "slug": "jack-morris",
        "name": "Jack Morris"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Jack Morris examines why general-purpose language models struggle with recent, private, and long-tail knowledge, then compares retrieval and agentic search with storing information in model weights. The workshop discusses embedding models, measured information capacity, catastrophic forgetting, supervised fine-tuning, training-data construction, and practical questions about maintaining specialized models."
  },
  {
    "slug": "substrate-launch-the-api-for-modular-ai",
    "title": "Substrate Launch: the API for modular AI",
    "url": "https://ai.engineer/talks/substrate-launch-the-api-for-modular-ai",
    "videoId": "x8HbIJh2wpQ",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 320000,
    "speakers": [
      {
        "slug": "rob-cheung",
        "name": "Rob Cheung",
        "organization": "Substrate"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Rob Cheung introduces Substrate, a modular AI platform combining a developer SDK for multimodal computation graphs with an inference engine optimized to execute those graphs efficiently. He contrasts multi-step AI systems with monolithic foundation models, describes nodes for image generation, speech transcription, text, JSON, embeddings, and code execution, and argues that integrated graph execution reduces per-step infrastructure overhead. He also highlights reliable, fast JSON decoding and the advantages of more legible, extensible AI applications."
  },
  {
    "slug": "supabase-vector-the-postgres-vector-database",
    "title": "Supabase Vector: The Postgres Vector database",
    "url": "https://ai.engineer/talks/supabase-vector-the-postgres-vector-database",
    "videoId": "MDxEXKkxf2Q",
    "event": "AI Engineer Summit 2023",
    "durationMs": 965000,
    "speakers": [
      {
        "slug": "paul-copplestone",
        "name": "Paul Copplestone",
        "organization": "Supabase"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Supabase co-founder Paul Copplestone presents PostgreSQL and pgvector as a practical foundation for embedding-powered AI applications. He outlines Supabase’s database, authentication, Deno-based edge functions, realtime features, and vector offering; discusses pgvector adoption; and compares IVFFlat with HNSW using throughput, accuracy, and comparisons involving Qdrant and Pinecone. The talk closes with PostgreSQL partitioning and potential work with Microsoft’s Citus team to support applications storing billions of embeddings."
  },
  {
    "slug": "supercharging-developer-workflow-with-amazon-q-developer",
    "title": "Supercharging Developer Workflow with Amazon Q Developer",
    "url": "https://ai.engineer/talks/supercharging-developer-workflow-with-amazon-q-developer",
    "videoId": "utxVvGJ4bcg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 802000,
    "speakers": [
      {
        "slug": "vikash-agrawal",
        "name": "Vikash Agrawal",
        "organization": "AWS"
      },
      {
        "slug": "linda",
        "name": "Linda",
        "organization": "AWS"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "AWS presenters Vikash Agrawal and Linda demonstrate Amazon Q Developer across the software development life cycle by generating a 2048 game with FastAPI through command-line and IDE workflows. They cover MCP, documentation, unit testing, context-aware bug fixes, and troubleshooting an AWS Lambda deployment with CloudWatch and Amazon Q Developer AI operations."
  },
  {
    "slug": "survive-the-ai-knife-fight-building-products-that-win",
    "title": "Survive the AI Knife-Fight: Building Products That Win",
    "url": "https://ai.engineer/talks/survive-the-ai-knife-fight-building-products-that-win",
    "videoId": "1MVh05GDydE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 850000,
    "speakers": [
      {
        "slug": "brian-balfour",
        "name": "Brian Balfour",
        "organization": "Reforge"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Reforge founder and CEO Brian Balfour explains how rapid AI product launches, incumbent platforms, and heavily funded startups make differentiation more difficult. He introduces Reforge's AI research and project-management agents, argues against both building unnecessary custom AI infrastructure and copying generic chatbot features, and recommends composing existing AI capabilities with proprietary product data, functionality, and clearly understood unmet customer needs."
  },
  {
    "slug": "swe-marathon-evaluating-coding-agents-at-billion-token-scale-rishi-desai-abundant-ai",
    "title": "SWE-Marathon: Evaluating Coding Agents at Billion-Token Scale - Rishi Desai, Abundant AI",
    "url": "https://ai.engineer/talks/swe-marathon-evaluating-coding-agents-at-billion-token-scale-rishi-desai-abundant-ai",
    "videoId": "Rx8f05JI_WA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 778000,
    "speakers": [
      {
        "slug": "rishi-desai",
        "name": "Rishi Desai",
        "organization": "Abundant AI"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Abundant AI engineer Rishi Desai presents SWE-Marathon, a 20-task benchmark measuring whether coding agents can complete project-scale software work over exceptionally long token and execution horizons. He explains its progression beyond earlier coding benchmarks, its combination of deterministic and browser-based computer-use verification for full-stack clones, its Harbor task format, and defenses against reward hacking. The recording reports a historical best resolution rate of 26%, an average trial of 31 million tokens, and a longest rollout of 877 million tokens; the public leaderboard has since changed."
  },
  {
    "slug": "swe-rebench-lessons-from-evaluating-coding-agents-on-real-software-engineering-tasks-ibragim-bad",
    "title": "SWE-rebench: Lessons from Evaluating Coding Agents on Real Software Engineering Tasks — Ibragim Badertdinov, Nebius",
    "url": "https://ai.engineer/talks/swe-rebench-lessons-from-evaluating-coding-agents-on-real-software-engineering-tasks-ibragim-bad",
    "videoId": "wcUJWP6WpGM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 990000,
    "speakers": [
      {
        "slug": "ibragim-badertdinov",
        "name": "Ibragim Badertdinov",
        "organization": "Nebius"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Nebius researcher Ibragim Badertdinov explains how SWE-rebench continuously evaluates coding agents on fresh, real-world GitHub software-engineering tasks. He describes containerized execution environments, regression testing, GitHub Archive ingestion, automated filtering and manual task validation, alongside practical concerns including agent-harness differences, token costs, caching and configuration drift. A Claude Code example demonstrates benchmark leakage: after direct web access was restricted, the agent retrieved the original GitHub issue using curl."
  },
  {
    "slug": "taking-reinforcement-learning-cross-datacenter",
    "title": "Taking Reinforcement Learning Cross Datacenter",
    "url": "https://ai.engineer/talks/taking-reinforcement-learning-cross-datacenter",
    "videoId": "maRzp4kImJ4",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1190000,
    "speakers": [
      {
        "slug": "nan-jiang",
        "name": "Nan Jiang",
        "organization": "Modal"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Modal engineer Nan Jiang describes how reinforcement-learning post-training can separate tightly coupled, RDMA-connected training from geographically distributed rollout workers. Rather than repeatedly transferring approximately 500 GB checkpoints, the approach exploits the observation that fewer than 1% of rollout-visible weights may change between versions because small Adam updates encounter finite-precision rounding boundaries. This enables sparse synchronization of served policy weights while distributed rollout islands return trajectories and metadata to a central trainer."
  },
  {
    "slug": "taming-rogue-ai-agents-with-observability-driven-evaluation",
    "title": "Taming Rogue AI Agents with Observability-Driven Evaluation",
    "url": "https://ai.engineer/talks/taming-rogue-ai-agents-with-observability-driven-evaluation",
    "videoId": "xJXm4Wcw4m8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 975000,
    "speakers": [
      {
        "slug": "jim-bennett",
        "name": "Jim Bennett",
        "organization": "Galileo"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Galileo principal developer advocate Jim Bennett explains why hallucinations, nondeterministic model behavior, and multistep agent workflows require observability-driven evaluation instead of conventional deterministic testing. Using a banking chatbot example, he shows how language models can score RAG outputs, identify missed tool calls, and diagnose low evaluation metrics, while specialized evaluation models and continuous human feedback help keep automated assessment accountable."
  },
  {
    "slug": "task-fidelity-scaling-laws-kobie-crawford-snorkel-ai",
    "title": "Task Fidelity Scaling Laws — Kobie Crawford, Snorkel AI",
    "url": "https://ai.engineer/talks/task-fidelity-scaling-laws-kobie-crawford-snorkel-ai",
    "videoId": "YYH0DMQr30A",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1240006,
    "speakers": [
      {
        "slug": "kobie-crawford",
        "name": "Kobie Crawford",
        "organization": "Snorkel AI"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Snorkel AI developer advocate Kobie Crawford explains how task fidelity influences agent evaluation and reinforcement-learning outcomes. Comparing accepted, well-specified tasks against rejected, lower-quality tasks, he argues that difficult tasks should produce meaningful model failures rather than noise from faulty specifications. In a controlled training comparison, lower-quality tasks improved the base model by about 1%, while higher-quality tasks improved it by about 6%. The presentation also discusses Harbor and OpenEnv, evaluations involving Sonnet 4.5 and Codex, and quality assurance using human experts, LLM judges, shared rubrics, and inter-annotator agreement; audience questions address task mixtures, instruction specificity, and longer-horizon challenges."
  },
  {
    "slug": "taste-craft-a-conversation-with-tuomas-artman-cto-of-linear-and-gergely-orosz-of-the-pragmatic-e",
    "title": "Taste & Craft: A Conversation with Tuomas Artman, CTO of Linear, and Gergely Orosz of The Pragmatic Engineer",
    "url": "https://ai.engineer/talks/taste-craft-a-conversation-with-tuomas-artman-cto-of-linear-and-gergely-orosz-of-the-pragmatic-e",
    "videoId": "wjk0ulMAkbc",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1757000,
    "speakers": [
      {
        "slug": "tuomas-artman",
        "name": "Tuomas Artman",
        "organization": "Linear"
      },
      {
        "slug": "gergely-orosz",
        "name": "Gergely Orosz",
        "organization": "The Pragmatic Engineer"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Linear cofounder and CTO Tuomas Artman and The Pragmatic Engineer's Gergely Orosz discuss why AI-assisted development makes product judgment, restraint, and craftsmanship more important. Drawing on their Uber experience, they examine feature overload, Claude Code and Claude Opus 4.5, Linear's reported automation of roughly 10% of bug fixes, and quality-focused engineering habits including weekly self-directed problem discovery."
  },
  {
    "slug": "teaching-ai-to-find-real-vulnerabilities-prof-david-brumley-bugcrowd",
    "title": "Teaching AI to Find Real Vulnerabilities — Prof. David Brumley, Bugcrowd",
    "url": "https://ai.engineer/talks/teaching-ai-to-find-real-vulnerabilities-prof-david-brumley-bugcrowd",
    "videoId": "ZFxh7sqbUZo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1637000,
    "speakers": [
      {
        "slug": "david-brumley",
        "name": "David Brumley",
        "organization": "Bugcrowd; Carnegie Mellon University"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "David Brumley, a Carnegie Mellon professor and Bugcrowd’s Chief AI and Science Officer, explains how to train and evaluate language models for genuine vulnerability discovery using reinforcement-learning environments modeled on progressive human cybersecurity training. He emphasizes realistic target difficulty, reproducible execution environments, deterministic grading instead of LLM judges, audit tasks that assess multiple vulnerabilities, and demanding V8 sandbox-escape scenarios involving exploit chains and arbitrary read/write primitives. He concludes by highlighting reward hacking and the challenge of evaluating newly discovered vulnerabilities."
  },
  {
    "slug": "teaching-coding-agents-to-do-spreadsheets",
    "title": "Teaching Coding Agents to do Spreadsheets",
    "url": "https://ai.engineer/talks/teaching-coding-agents-to-do-spreadsheets",
    "videoId": "HEFSExa0xl0",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1149000,
    "speakers": [
      {
        "slug": "nuno-campos",
        "name": "Nuno Campos",
        "organization": "Witan Labs"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Nuno Campos explains how Witan Labs improved an internal financial-spreadsheet benchmark from roughly 50% to 92%. After finding multi-agent orchestration and standalone SQL, XML, CSV/TSV, and HTML representations insufficient, the team combined spreadsheet rendering and a high-fidelity C# engine with a persistent Node.js REPL. Unlike stateless Code Mode, the REPL preserves variables between calls, supports shorter JavaScript programs interleaved with model reasoning, and consolidates spreadsheet tools. Better search, formula tracing, systematic evaluation, and debugging supplied additional improvements."
  },
  {
    "slug": "teaching-gemini-to-speak-youtube-adapting-llms-for-video-recommendations-to-2b-dau",
    "title": "Teaching Gemini to Speak YouTube: Adapting LLMs for Video Recommendations to 2B+ DAU",
    "url": "https://ai.engineer/talks/teaching-gemini-to-speak-youtube-adapting-llms-for-video-recommendations-to-2b-dau",
    "videoId": "LxQsQ3vZDqo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1371000,
    "speakers": [
      {
        "slug": "devansh-tandon",
        "name": "Devansh Tandon",
        "organization": "Google / YouTube"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Devansh Tandon explains how YouTube adapts Gemini checkpoints into a Large Recommender Model for recommendation retrieval and ranking. Videos are represented by Semantic IDs derived from multimodal content embeddings and quantized with RQ-VAE, giving the model a domain-specific token language that can be aligned with natural language. He illustrates personalized generative retrieval, production deployment, and a reusable adaptation recipe before audience questions address cold starts, visual encoders, and pre-training data."
  },
  {
    "slug": "text-diffusion-brendan-o-donoghue-google-deepmind",
    "title": "Text Diffusion — Brendan O'Donoghue, Google DeepMind",
    "url": "https://ai.engineer/talks/text-diffusion-brendan-o-donoghue-google-deepmind",
    "videoId": "r305-aQTaU0",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1683000,
    "speakers": [
      {
        "slug": "brendan-o-donoghue",
        "name": "Brendan O'Donoghue",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Google DeepMind researcher Brendan O'Donoghue explains how Gemini Diffusion generates text by iteratively denoising corrupted token sequences rather than predicting tokens strictly one at a time. He contrasts diffusion and autoregressive models, discussing accelerator utilization, latency and throughput tradeoffs, bidirectional reasoning, self-correction, adaptive computation, and GPQA Diamond evaluation. A voice-driven coding demonstration illustrates potential low-latency applications, followed by audience questions about training, reinforcement learning, scaling, pricing, and hybrid generation."
  },
  {
    "slug": "text-to-speech-data-preparation-and-fine-tuning-workshop-ronan-mcgovern",
    "title": "Text-to-Speech Data Preparation and Fine-tuning Workshop - Ronan McGovern",
    "url": "https://ai.engineer/talks/text-to-speech-data-preparation-and-fine-tuning-workshop-ronan-mcgovern",
    "videoId": "CXsbjcrf_5g",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2040000,
    "speakers": [
      {
        "slug": "ronan-mcgovern",
        "name": "Ronan McGovern",
        "organization": "Trelis Research"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Ronan McGovern of Trelis Research demonstrates an end-to-end online workshop for adapting Sesame's CSM-1B text-to-speech model to a target voice. He explains its hierarchical, two-transformer audio-token architecture, prepares a speech dataset from YouTube audio using Whisper, runs a Google Colab fine-tuning workflow with Unsloth and adapters, distinguishes voice cloning from fine-tuning, and evaluates generated speech before and after training. The accompanying notebook and slides are available in the TrelisResearch/ai-worlds-fair-2025 repository."
  },
  {
    "slug": "the-1-000x-ai-engineer-swyx",
    "title": "The 1,000x AI Engineer: Swyx",
    "url": "https://ai.engineer/talks/the-1-000x-ai-engineer-swyx",
    "videoId": "qaJXBMwUkoE",
    "event": "AI Engineer Summit 2023",
    "durationMs": 567000,
    "speakers": [
      {
        "slug": "swyx",
        "name": "swyx",
        "organization": "Latent.Space"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Swyx frames AI engineering as an emerging software discipline created by rapidly advancing models and a developer population much larger than the specialized machine-learning workforce. Drawing on historical technology cycles, AlexNet's 2012 breakthrough, and rising model-training compute, he distinguishes three categories: engineers enhanced by AI tools, engineers building AI products, and nonhuman AI engineer agents, illustrated by Copilot, Midjourney, and AutoGPT."
  },
  {
    "slug": "the-100-tool-agent-is-a-trap-scaling-with-semantic-routers-and-jit-context",
    "title": "The 100-Tool Agent Is a Trap: Scaling with Semantic Routers and JIT Context",
    "url": "https://ai.engineer/talks/the-100-tool-agent-is-a-trap-scaling-with-semantic-routers-and-jit-context",
    "videoId": "vh2VGuQ3zhY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1707000,
    "speakers": [
      {
        "slug": "sohail-shaikh",
        "name": "Sohail Shaikh",
        "organization": "Prosodica"
      },
      {
        "slug": "ankush-rastogi",
        "name": "Ankush Rastogi",
        "organization": "Prosodica"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Prosodica’s Sohail Shaikh and Ankush Rastogi explain why a “Fat Agent” that includes every tool schema in every prompt becomes slower, more expensive, and less accurate as its tool catalog grows. They present semantic routing as retrieval-augmented generation for tools, combining embeddings and vector search with just-in-time context injection so each request receives only relevant capabilities. The talk examines time-to-first-token latency, tool-selection accuracy, Anthropic’s documented MCP token-reduction example, and evaluating top-K retrieval settings against production accuracy, cost, and latency targets."
  },
  {
    "slug": "the-2025-ai-engineering-report-barr-yaron-amplify-partners",
    "title": "The 2025 AI Engineering Report — Barr Yaron, Amplify Partners",
    "url": "https://ai.engineer/talks/the-2025-ai-engineering-report-barr-yaron-amplify-partners",
    "videoId": "mQ7_Zje7WKE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 753000,
    "speakers": [
      {
        "slug": "barr-yaron",
        "name": "Barr Yaron",
        "organization": "Amplify Partners"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Amplify Partners investment partner Barr Yaron presents findings from the 2025 State of AI Engineering Survey, covering practitioners' varied job titles and AI experience, widespread internal and customer-facing LLM deployments, and OpenAI model adoption. She reports that 70% of respondents use RAG, discusses LoRA, QLoRA, DPO, and supervised fine-tuning, and notes that 70% update prompts at least monthly while 31% lack structured prompt management. The talk also addresses the multimodal production gap, AI agents, and model-usage monitoring."
  },
  {
    "slug": "the-2026-state-of-ai-engineering-barr-yaron-amplify-partners",
    "title": "The 2026 State of AI Engineering — Barr Yaron, Amplify Partners",
    "url": "https://ai.engineer/talks/the-2026-state-of-ai-engineering-barr-yaron-amplify-partners",
    "videoId": "RGe6EjucbzI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1187000,
    "speakers": [
      {
        "slug": "barr-yaron",
        "name": "Barr Yaron",
        "organization": "Amplify Partners"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Amplify Partners partner Barr Yaron presents the 2026 AI engineering survey, conducted with Notion and Vercel, examining practitioners’ experience, modality adoption, model selection, infrastructure costs, agent controls, and team dynamics. She highlights rapid interest in audio and image generation, widespread combined use of open-weight and closed models, increasing tool standardization, production monitoring of token costs, and reliance on human approvals and permission gating for agents."
  },
  {
    "slug": "the-3-pillars-of-autonomy-michele-catasta-replit",
    "title": "The 3 Pillars of Autonomy – Michele Catasta, Replit",
    "url": "https://ai.engineer/talks/the-3-pillars-of-autonomy-michele-catasta-replit",
    "videoId": "MLhAA9yguwM",
    "event": "AI Engineer Code 2025",
    "durationMs": 1482000,
    "speakers": [
      {
        "slug": "michele-catasta",
        "name": "Michele Catasta",
        "organization": "Replit"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Replit's Michele Catasta explains why coding agents for nontechnical users must complete substantial software-development tasks without expert supervision. He contrasts supervised and fully autonomous experiences, traces the progression toward native tool calling and stronger frontier models, and defines reducible runtime as uninterrupted agent execution without technical user decisions. He then examines automated verification across static analysis, unit and integration tests, API checks, browser interactions, and directly generated Playwright code, alongside challenges in maintaining coherent long-running agent loops."
  },
  {
    "slug": "the-4-patterns-of-ai-native-development",
    "title": "The 4 Patterns of AI Native Development",
    "url": "https://ai.engineer/talks/the-4-patterns-of-ai-native-development",
    "videoId": "9u6xvcNJaxc",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 851000,
    "speakers": [
      {
        "slug": "patrick-debois",
        "name": "Patrick Debois"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Patrick Debois outlines four shifts in AI-native software development: developers move from producing code to managing coding agents, from implementation details to specifying intent, from delivery toward experimentation and discovery, and from creating content to capturing reusable organizational knowledge. Examples address increased code-review demands and cognitive load, agent permissions and guardrails, Markdown specifications, GitHub task planning, customer-led prototyping, onboarding, and preserving feature decisions."
  },
  {
    "slug": "the-a-to-z-of-building-ai-agents",
    "title": "The A to Z of Building AI Agents",
    "url": "https://ai.engineer/talks/the-a-to-z-of-building-ai-agents",
    "videoId": "IRp7lvBlbHs",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1653000,
    "speakers": [
      {
        "slug": "apoorva-joshi",
        "name": "Apoorva Joshi",
        "organization": "MongoDB"
      },
      {
        "slug": "ben-perlmutter",
        "name": "Ben Perlmutter",
        "organization": "MongoDB"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "MongoDB developer advocate Apoorva Joshi introduces the scheduled two-instructor workshop on building an AI research agent. The recording explains when agents are appropriate, LLM planning and tools, ReAct and reflection, application-specific long-term memory, LangChain function calling, arXiv paper retrieval and summarization, and wrapping Python functions with RunnableLambda before hands-on setup and exercises."
  },
  {
    "slug": "the-adversarial-path-to-the-personal-assistant",
    "title": "The Adversarial Path to the Personal Assistant",
    "url": "https://ai.engineer/talks/the-adversarial-path-to-the-personal-assistant",
    "videoId": "ckLXb15BnM8",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1126000,
    "speakers": [
      {
        "slug": "sumit-agarwal",
        "name": "Sumit Agarwal",
        "organization": "Ario"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Ario’s Sumit Agarwal presents a personalized AI assistant built on adversarial ETL: collecting a user’s existing activity across services such as Google, Amazon, DoorDash, and Strava, then combining that data with RAG, LLMs, and an agent architecture. Demonstrations cover personalized data portraits and recommendations, avoiding repetitive onboarding questions, and assigning straightforward computations to deterministic systems instead of language models. The talk also references Ario’s iOS application and Ario Boost."
  },
  {
    "slug": "the-age-of-the-agent",
    "title": "The Age of the Agent",
    "url": "https://ai.engineer/talks/the-age-of-the-agent",
    "videoId": "lFMKXnpbhpQ",
    "event": "AI Engineer Summit 2023",
    "durationMs": 889000,
    "speakers": [
      {
        "slug": "flo-crivello",
        "name": "Flo Crivello",
        "organization": "Lindy"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Lindy founder Flo Crivello argues that agentic AI could democratize business creation much as YouTube democratized media. Drawing on his experience founding Teamflow, he contrasts generative content tools with assistants that perform administrative work, demonstrates meeting coordination, and describes cooperating Lindy agents as a way for small teams or individuals to achieve substantially greater impact."
  },
  {
    "slug": "the-agent-awakens-collaborative-development-with-copilot",
    "title": "The Agent Awakens: Collaborative Development with Copilot",
    "url": "https://ai.engineer/talks/the-agent-awakens-collaborative-development-with-copilot",
    "videoId": "tHJSZ1-ZqcA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 3846000,
    "speakers": [
      {
        "slug": "christopher-harrison",
        "name": "Christopher Harrison",
        "organization": "GitHub"
      },
      {
        "slug": "john-peck",
        "name": "John Peck",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "GitHub enterprise advocate Christopher Harrison leads an interactive GitHub Copilot lab, with John Peck assisting participants. The workshop covers account and repository setup, treating Copilot as an AI pair programmer, giving coding agents useful context, security and agent fallibility, repository and path-specific instruction files, and extending Copilot Chat and coding agents with MCP servers, including enterprise integration and namespace questions."
  },
  {
    "slug": "the-agent-development-life-cycle",
    "title": "The Agent Development Life Cycle",
    "url": "https://ai.engineer/talks/the-agent-development-life-cycle",
    "videoId": "0vBKv9yAQi4",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1120000,
    "speakers": [
      {
        "slug": "zack-reneau-wedeen",
        "name": "Zack Reneau-Wedeen",
        "organization": "Sierra"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Sierra product manager Zack Reneau-Wedeen connects his early work on Google Lens and the nondeterminism of AI systems to Sierra's methodology for building reliable conversational agents. He discusses production-grade developer and customer-experience platforms, conversation-level quality assurance and reporting, practical voice-agent testing with Deepgram and Mac audio-routing tools, and the potential of voice-to-voice models."
  },
  {
    "slug": "the-agent-native-company",
    "title": "The Agent-Native Company",
    "url": "https://ai.engineer/talks/the-agent-native-company",
    "videoId": "0ZPAvzhpGjw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1258000,
    "speakers": [
      {
        "slug": "rick-blalock",
        "name": "Rick Blalock",
        "organization": "Agentuity"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Agentuity founder Rick Blalock distinguishes agent-native companies from organizations that merely add AI tools, arguing that agents must be foundational to products, operations, and culture. Using examples including Devin, automated GitHub changelogs and documentation, content-marketing agents, rapid prototyping, and emerging agent-management roles, he describes humans orchestrating AI coworkers and explains how agent fluency can reshape hiring and organizational structure."
  },
  {
    "slug": "the-agent-ready-web-simplify-user-actions-with-webmcp",
    "title": "The agent-ready web: Simplify user actions with WebMCP",
    "url": "https://ai.engineer/talks/the-agent-ready-web-simplify-user-actions-with-webmcp",
    "videoId": "ghJmWQCIHRM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1294000,
    "speakers": [
      {
        "slug": "tara-agyemang",
        "name": "Tara Agyemang",
        "organization": "Google"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Google Chrome developer relations engineer Tara Agyemang introduces WebMCP, a proposed web standard that exposes structured website tools to AI agents instead of relying on fragile DOM inspection, screenshots, and screen scraping. She explains why semantic HTML, accessibility, performance, and Core Web Vitals remain prerequisites for agent-ready websites, then demonstrates Gemini-powered maze-game and concert-ticket workflows using the Model Context Tool Inspector. Google’s primary documentation corroborates declarative HTML-form and imperative JavaScript implementation paths."
  },
  {
    "slug": "the-agentic-ai-engineer",
    "title": "The Agentic AI Engineer",
    "url": "https://ai.engineer/talks/the-agentic-ai-engineer",
    "videoId": "pSto5YaNGUo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2090000,
    "speakers": [
      {
        "slug": "benedikt-sanftl",
        "name": "Benedikt Sanftl",
        "organization": "Mutagent"
      },
      {
        "slug": "burak-cemil-ozafsar",
        "name": "Burak Cemil Özafşar",
        "organization": "Mutagent"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Mutagent co-founders Benedikt Sanftl and Burak Cemil Özafşar present an agentic AI engineering workflow that combines offline build-and-evaluate loops with online monitoring, trace diagnosis, and optimization. They discuss agent specifications, evaluation datasets and metrics, calibrated LLM-as-a-judge scoring, harness effects, and orchestration across tools such as Claude Code, Codex, Langfuse, and GitHub, followed by a product demonstration."
  },
  {
    "slug": "the-agentic-web-and-the-bazaar-era-of-ai-ramesh-raskar-mit-media-lab",
    "title": "The Agentic Web and the Bazaar Era of AI - Ramesh Raskar, MIT Media Lab",
    "url": "https://ai.engineer/talks/the-agentic-web-and-the-bazaar-era-of-ai-ramesh-raskar-mit-media-lab",
    "videoId": "sum9DgexFRQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 731000,
    "speakers": [
      {
        "slug": "ramesh-raskar",
        "name": "Ramesh Raskar",
        "organization": "MIT Media Lab"
      },
      {
        "slug": "maria-gorskikh",
        "name": "Maria Gorskikh",
        "organization": "Project NANDA"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Ramesh Raskar and Project NANDA contributor Maria describe open infrastructure for an agentic web where AI agents can discover one another, establish trustworthy identities, transact, and coordinate across organizational boundaries. They explain the NANDA Index, signed AgentFacts, agent cards, Host39 onboarding, DNS-AID, OpenClaw, Maritime’s cost-conscious agent hosting, and NandaTown simulation, framing the ecosystem around discovery, commerce, and a collaborative bazaar."
  },
  {
    "slug": "the-ai-bugpocalypse-is-here-now-what",
    "title": "The AI bugpocalypse is here. Now what?",
    "url": "https://ai.engineer/talks/the-ai-bugpocalypse-is-here-now-what",
    "videoId": "7JgIS42mz7U",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1184000,
    "speakers": [
      {
        "slug": "jack-cable",
        "name": "Jack Cable",
        "organization": "Corridor"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Corridor co-founder and CEO Jack Cable explains how increasingly capable frontier models and autonomous coding agents simultaneously accelerate vulnerability discovery and expand software attack surfaces. Drawing on Secure by Design, MITRE vulnerability classes, memory-safe languages, and the BaxBench benchmark, he argues for resilient-by-design software, security guardrails before pull requests, visibility into AI coding tools, and a competitive ecosystem of open-weight models."
  },
  {
    "slug": "the-ai-emperor-has-no-daus-why-most-devs-still-don-t-use-code-ai",
    "title": "The AI emperor has no DAUs: why most devs still don't use code AI",
    "url": "https://ai.engineer/talks/the-ai-emperor-has-no-daus-why-most-devs-still-don-t-use-code-ai",
    "videoId": "Up6WVA07QdE",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1125000,
    "speakers": [
      {
        "slug": "quinn-slack",
        "name": "Quinn Slack",
        "organization": "Sourcegraph"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Sourcegraph CEO and cofounder Quinn Slack argues that actual, recurring use of AI coding tools remains much lower than headline survey figures suggest. He scrutinizes GitHub survey scope, paid Copilot subscriptions, and active-user metrics, presenting his roughly 5% adoption figure as a personal estimate rather than an established industry fact. Drawing on Sourcegraph's Cody, he discusses developer hesitation, the limits of autocomplete and ghost text, and product-design principles for delivering readily verifiable value and sustained usage."
  },
  {
    "slug": "the-ai-engineer-s-guide-to-raising-vc-dani-grant-jam-chelcie-taylor-notable-capital",
    "title": "The AI Engineer’s Guide to Raising VC — Dani Grant (Jam), Chelcie Taylor (Notable Capital)",
    "url": "https://ai.engineer/talks/the-ai-engineer-s-guide-to-raising-vc-dani-grant-jam-chelcie-taylor-notable-capital",
    "videoId": "YYNXFsUutbM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2057000,
    "speakers": [
      {
        "slug": "dani-grant",
        "name": "Dani Grant",
        "organization": "Jam"
      },
      {
        "slug": "chelcie-taylor",
        "name": "Chelcie Taylor",
        "organization": "Notable Capital"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Dani Grant of Jam and Chelcie Taylor of Notable Capital present an interactive fundraising workshop for AI startup founders. They explain why pre-seed investors may invest before revenue, a finished product, or a full-time founding team; discuss founder vision, co-founders, investor psychology, warm introductions and cold outreach; and cover differentiated AI pitches, voice agents, ideal customer profiles, and team evaluation."
  },
  {
    "slug": "the-ai-evolution",
    "title": "The AI Evolution",
    "url": "https://ai.engineer/talks/the-ai-evolution",
    "videoId": "yLa3i3rAdEE",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1172000,
    "speakers": [
      {
        "slug": "mario-rodriguez",
        "name": "Mario Rodriguez",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "GitHub product executive Mario Rodriguez traces GitHub Copilot from internal 2020 exploration and early unsuccessful pull-request experiments to its 2021 technical preview. He discusses adoption, code-completion usage, and reported annual recurring revenue, then explains why developer-flow integration, low latency, offline evaluations, and quality scorecards matter for AI products. He concludes by describing deeper codebase understanding and AI integration across GitHub's broader development platform, including Actions and Codespaces."
  },
  {
    "slug": "the-ai-pivot-with-chris-white-of-prefect-bryan-bischof-of-hex",
    "title": "The AI Pivot: With Chris White of Prefect & Bryan Bischof of Hex",
    "url": "https://ai.engineer/talks/the-ai-pivot-with-chris-white-of-prefect-bryan-bischof-of-hex",
    "videoId": "bDQn1w0idXs",
    "event": "AI Engineer Summit 2023",
    "durationMs": 2116000,
    "speakers": [
      {
        "slug": "chris-white",
        "name": "Chris White",
        "organization": "Prefect"
      },
      {
        "slug": "bryan-bischof",
        "name": "Bryan Bischof",
        "organization": "Hex"
      },
      {
        "slug": "brittany-walker",
        "name": "Brittany Walker",
        "organization": "CRV"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "CRV principal Brittany Walker moderates a discussion with Prefect CTO Chris White and Hex AI lead Bryan Bischof about how established software companies decide to invest in AI and integrate it into existing products. They discuss orchestration for unreliable model APIs, semantic search, retrieval, ETL and vector databases; Prefect's Marvin project; Hex Magic and text-to-SQL product design; hiring, rollout strategy, model evaluation, and measuring product-level returns."
  },
  {
    "slug": "the-art-science-of-benchmarking-agents",
    "title": "The Art & Science of Benchmarking Agents",
    "url": "https://ai.engineer/talks/the-art-science-of-benchmarking-agents",
    "videoId": "iNkFlCiij0U",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1405000,
    "speakers": [
      {
        "slug": "vincent-chen",
        "name": "Vincent Chen",
        "organization": "Snorkel AI"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Snorkel AI co-founder Vincent Chen outlines how agent benchmarks can close the widening gap between AI capabilities and evaluation. Drawing on Snorkel's $3 million Open Benchmarks Grants, he distinguishes scientific requirements such as expert-validated tasks, adversarial quality control, broad task coverage, and meaningful headroom from practical concerns including researcher adoption, policy adherence, and long-horizon organizational context. Examples include GPQA, MMLU, Terminal-Bench, and ARC-AGI-3, whose launch exposed a sharp human-versus-frontier-model performance gap."
  },
  {
    "slug": "the-base-model-is-dead",
    "title": "The Base Model is Dead",
    "url": "https://ai.engineer/talks/the-base-model-is-dead",
    "videoId": "xbPriQWXtWM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1065000,
    "speakers": [
      {
        "slug": "varun-singh",
        "name": "Varun Singh",
        "organization": "Arcee AI"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Arcee AI pre-training lead Varun Singh argues that modern base models no longer primarily mirror scraped web text: their training increasingly emphasizes code, STEM, reasoning, and data useful for downstream reinforcement learning. Comparing GPT-3, Llama 3, Arcee's Trinity-Large-Thinking, MAI-Thinking-1, and Nemotron 3 Ultra, he contrasts synthetic-data strategies and describes how supervised fine-tuning-style examples and post-training requirements are moving earlier into pre-training."
  },
  {
    "slug": "the-benchmarks-game-why-it-s-rigged-and-how-you-can-really-win",
    "title": "The Benchmarks Game: Why It's Rigged and How You Can (Really) Win",
    "url": "https://ai.engineer/talks/the-benchmarks-game-why-it-s-rigged-and-how-you-can-really-win",
    "videoId": "EnT4Wej5M5k",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 680000,
    "speakers": [
      {
        "slug": "darius-emrani",
        "name": "Darius Emrani",
        "organization": "Scorecard"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Scorecard CEO Darius Emrani argues that AI benchmark incentives distort model comparisons through selective inference configurations, privileged benchmark access, and preference for polished responses over accuracy. Using Grok 3, FrontierMath, Llama 4 Maverick, SWE-bench, and Goodhart's law as examples, he advocates evaluations grounded in the actual deployment domain and task. Primary sources correct the transcript's AutoRover reference to AutoCodeRover and qualify its claim that OpenAI accessed the entire FrontierMath dataset: Epoch AI describes a separate holdout."
  },
  {
    "slug": "the-billable-hour-is-dead-long-live-the-billable-hour",
    "title": "The Billable Hour is Dead; Long Live the Billable Hour?",
    "url": "https://ai.engineer/talks/the-billable-hour-is-dead-long-live-the-billable-hour",
    "videoId": "Wv1tAxKYLeE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1024000,
    "speakers": [
      {
        "slug": "kevin-madura",
        "name": "Kevin Madura",
        "organization": "AlixPartners"
      },
      {
        "slug": "mo-bhasin",
        "name": "Mo Bhasin",
        "organization": "AlixPartners"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Kevin Madura and Mo Bhasin explain how AlixPartners applies an internal generative-AI platform to professional-services work, contrasting enthusiasm for enterprise AI with uneven realized productivity. They describe compressing document-ingestion and analysis workflows, examining entire contract corpora, classifying business records with NAICS codes, extracting structured information from credit agreements, and pairing technical deployments with organizational adoption."
  },
  {
    "slug": "the-bitter-layout-or-how-i-learned-to-love-the-model-picker",
    "title": "The Bitter Layout or: How I Learned to Love the Model Picker",
    "url": "https://ai.engineer/talks/the-bitter-layout-or-how-i-learned-to-love-the-model-picker",
    "videoId": "BZtD0yYAgCQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 864000,
    "speakers": [
      {
        "slug": "maximillian-piras",
        "name": "Maximillian Piras",
        "organization": "Yutori"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Yutori designer Maximillian Piras examines why AI applications converge on the Bitter Layout: a chat input, turn-by-turn interaction, and a model picker. He contrasts persistent criticism of chat usability with widespread adoption, interprets model selectors as troublesome interface modes, and argues that conversational interfaces endure because their flexibility lets them adapt to rapidly changing models. The talk concludes by encouraging designers to rethink AI interfaces around evolving capabilities and higher-level design goals."
  },
  {
    "slug": "the-build-operate-divide-bridging-product-vision-and-ai-operational-reality",
    "title": "The Build-Operate Divide: Bridging Product Vision and AI Operational Reality",
    "url": "https://ai.engineer/talks/the-build-operate-divide-bridging-product-vision-and-ai-operational-reality",
    "videoId": "1__V4KTv_Gw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 770000,
    "speakers": [
      {
        "slug": "jeremy-silva",
        "name": "Jeremy Silva",
        "organization": "Freeplay"
      },
      {
        "slug": "chris-hernandez",
        "name": "Chris Hernandez",
        "organization": "Chime"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Freeplay product leader Jeremy Silva and Chime speech-analytics leader Chris Hernandez explain why promising generative AI prototypes encounter operational and reliability problems in production. They advocate iterative evaluations, human-in-the-loop review, hallucination monitoring, and feedback loops, while positioning customer-experience and quality-assurance teams as prompt testers, model reviewers, and ongoing AI performance monitors."
  },
  {
    "slug": "the-code-ai-maturity-model-and-what-it-means-for-you",
    "title": "The Code AI Maturity Model and What It Means For You",
    "url": "https://ai.engineer/talks/the-code-ai-maturity-model-and-what-it-means-for-you",
    "videoId": "jpKCIVlS9wM",
    "event": "AI Engineer Summit 2023",
    "durationMs": 478000,
    "speakers": [
      {
        "slug": "ado-kukic",
        "name": "Ado Kukic",
        "organization": "Sourcegraph"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Sourcegraph’s Ado Kukic presents a six-level Code AI maturity model, modeled on SAE vehicle-autonomy levels, progressing from manual coding through autocomplete, codebase-aware assistance, requirements-driven feature generation, autonomous pull requests, and ultimately end-to-end AI-led software development. He illustrates the framework with Sourcegraph’s Cody, Node.js HTTP-client selection, automated documentation and bug fixes, and Waymo’s autonomous-driving analogy."
  },
  {
    "slug": "the-coherence-trap-why-llms-feel-smart-but-aren-t-thinking",
    "title": "The Coherence Trap: Why LLMs Feel Smart (But Aren’t Thinking)",
    "url": "https://ai.engineer/talks/the-coherence-trap-why-llms-feel-smart-but-aren-t-thinking",
    "videoId": "u825uxb7LnA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1247000,
    "speakers": [
      {
        "slug": "travis-frisinger",
        "name": "Travis Frisinger",
        "organization": "8th Light"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Travis Frisinger, Technical Director of AI at 8th Light, argues that the apparent intelligence of large language models is better understood as coherent, context-sensitive output than genuine understanding. Drawing on GPT-4, Microsoft Research's AGI discussion, live ChatGPT-assisted programming, the WebCat Azure Function, and creative multimodal experiments, he develops a coherence framework for conceptual activation during inference and recommends designing human-AI workflows around structured resonance."
  },
  {
    "slug": "the-cure-for-the-vibe-coding-hangover",
    "title": "The Cure for the Vibe Coding Hangover",
    "url": "https://ai.engineer/talks/the-cure-for-the-vibe-coding-hangover",
    "videoId": "JsKTQbT58BY",
    "event": "AI Engineer Code 2025",
    "durationMs": 3422000,
    "speakers": [
      {
        "slug": "corey-j-gallon",
        "name": "Corey J. Gallon",
        "organization": "Rexmore"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Corey J. Gallon presents a principles-process-tools framework for replacing brittle, unplanned vibe coding with maintainable AI-assisted software development. He covers project specifications, validation contracts, sequential atomic feature implementation, curated context packages, isolated execution, multisensory feedback, declarative document linking, and Git-backed progress tracking."
  },
  {
    "slug": "the-current-state-of-browser-agents",
    "title": "The Current State of Browser Agents",
    "url": "https://ai.engineer/talks/the-current-state-of-browser-agents",
    "videoId": "Djv8Sp11UjI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1273000,
    "speakers": [
      {
        "slug": "jerry-wu",
        "name": "Jerry Wu",
        "organization": "Halluminate"
      },
      {
        "slug": "wyatt-marshall",
        "name": "Wyatt Marshall",
        "organization": "Halluminate"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Halluminate founders Jerry Wu and Wyatt Marshall explain how browser agents operate, then use WebBench to assess real-world information-retrieval and action-taking workflows. They contrast screenshot-based and HTML/DOM-based approaches, discuss human-supervised baselines, and show how multistep tasks, authentication, anti-bot protections, CAPTCHAs, latency, and browser infrastructure affect reliability. Their practical guidance emphasizes selecting suitable use cases and infrastructure before deploying browser agents."
  },
  {
    "slug": "the-dark-arts-of-web-automation-teaching-agents-to-use-websites-like-humans",
    "title": "The Dark Arts of Web Automation: Teaching Agents to Use Websites Like Humans",
    "url": "https://ai.engineer/talks/the-dark-arts-of-web-automation-teaching-agents-to-use-websites-like-humans",
    "videoId": "26RtyAm9y_Q",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1298000,
    "speakers": [
      {
        "slug": "corey-gallon",
        "name": "Corey Gallon",
        "organization": "Rexmore"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Corey Gallon of Rexmore presents chrome-agent, an open-source CLI for controlling Chrome through the Chrome DevTools Protocol, and argues that reusable command-line workflows can reduce the model round trips associated with MCP-based browser automation. He describes a sense-act-verify loop and a three-rung interaction ladder, escalating from basic automation to trusted browser input and humanlike cursor movement. Demonstrations cover enterprise web interfaces, Cloudflare Turnstile, isolated shadow roots and cross-origin iframes, image CAPTCHAs, and reCAPTCHA v2; the speaker states that demonstrations use infrastructure and accounts he controls."
  },
  {
    "slug": "the-demo-i-wish-i-d-had-openai-s-agents-sdk-serverless",
    "title": "The Demo I Wish I'd Had: OpenAI's Agents SDK... serverless!",
    "url": "https://ai.engineer/talks/the-demo-i-wish-i-d-had-openai-s-agents-sdk-serverless",
    "videoId": "UcW_s4BmuD0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1071000,
    "speakers": [
      {
        "slug": "brook-riggio",
        "name": "Brook Riggio",
        "organization": "thryv.com"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Brook Riggio demonstrates a practical serverless AI application architecture that combines a Next.js frontend, the OpenAI Agents SDK, Inngest orchestration, and Python functions deployed on Vercel. He explains event-driven workflows, automatic step retries, tracing, observability, preview deployments, and a newsletter-generation demo that turns user-supplied topics into an orchestrated agent workflow."
  },
  {
    "slug": "the-desktop-frontier-ahmad-osman-osmantic",
    "title": "The Desktop Frontier — Ahmad Osman, Osmantic",
    "url": "https://ai.engineer/talks/the-desktop-frontier-ahmad-osman-osmantic",
    "videoId": "XV2oYi7kojc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1082000,
    "speakers": [
      {
        "slug": "ahmad-osman",
        "name": "Ahmad Osman",
        "organization": "Osmantic"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Osmantic founder Ahmad Osman argues that rapid improvements in open-model capability density are moving frontier-class AI from data centers onto desktop workstations and consumer GPUs. He compares successive Llama, Qwen, GLM, and Nemotron models, connects declining hardware requirements to published research on the densing law, and advocates sovereign AI, enterprise ownership of local compute, and the increasing practical utility of existing GPU hardware."
  },
  {
    "slug": "the-devops-engineer-who-never-sleeps",
    "title": "The Devops Engineer Who Never Sleeps",
    "url": "https://ai.engineer/talks/the-devops-engineer-who-never-sleeps",
    "videoId": "VZzUhELgYk4",
    "event": "AI Engineer Summit 2025",
    "durationMs": 978000,
    "speakers": [
      {
        "slug": "diamond-bishop",
        "name": "Diamond Bishop",
        "organization": "Datadog"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Datadog's Diamond Bishop presents Bits AI agents designed to automate DevOps and software engineering work. An AI On-Call Engineer investigates alerts using runbooks, logs, metrics, traces, hypotheses, and tool-based reasoning, while an AI Software Engineer identifies application errors and can generate fixes for GitHub or VS Code. The talk also discusses human-agent collaboration, evaluation challenges, AI-oriented team practices, and full-stack observability including GPUs and LLM monitoring."
  },
  {
    "slug": "the-dirty-secret-of-forward-deployed-engineering",
    "title": "The Dirty Secret of Forward Deployed Engineering",
    "url": "https://ai.engineer/talks/the-dirty-secret-of-forward-deployed-engineering",
    "videoId": "Byv311hdoHE",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1009000,
    "speakers": [
      {
        "slug": "natalie-meurer",
        "name": "Natalie Meurer",
        "organization": "Sierra"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Sierra's Natalie Meurer traces forward-deployed engineering from Palantir's on-site deployments and DevOps firefighting through customer data integration, Ontology modeling, platform enablement, and modern AI-agent work. She argues that the title encompasses several historical role vintages rather than one consistent job, while customer accountability and measurable outcomes remain its durable core as coding agents make software production cheaper."
  },
  {
    "slug": "the-emerging-skillset-of-wielding-coding-agents",
    "title": "The emerging skillset of wielding coding agents",
    "url": "https://ai.engineer/talks/the-emerging-skillset-of-wielding-coding-agents",
    "videoId": "F_RyElT_gJk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 2106000,
    "speakers": [
      {
        "slug": "beyang-liu",
        "name": "Beyang Liu",
        "organization": "Sourcegraph/Amp"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Sourcegraph co-founder and then-CTO Beyang Liu explains why coding agents polarize developers and argues that effective use requires abandoning mental models inherited from autocomplete and chatbots. He traces the shift from GPT-3 completion and GPT-3.5 chat interfaces to autonomous agents, discusses inference-cost incentives, and demonstrates agent planning, parallel work, VS Code diff review, and MCP configuration. Practical recommendations emphasize fast Playwright-powered feedback loops, codebase understanding, onboarding, and more effective code review."
  },
  {
    "slug": "the-end-of-apps",
    "title": "The End of Apps",
    "url": "https://ai.engineer/talks/the-end-of-apps",
    "videoId": "4fntwuOoedA",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1218000,
    "speakers": [
      {
        "slug": "kitze",
        "name": "Kitze",
        "organization": "Sizzy"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Kitze traces his progression from text-file task management, Tasker, IFTTT, and Google Assistant to Benji, a personal life operating system combining tasks, habits, and calendars. He explains how ChatGPT plugins, structured model output, voice-triggered APIs, and OpenClaw accelerated his shift toward self-hosted, purpose-specific personal agents, arguing that agent-mediated interactions will increasingly replace conventional app interfaces."
  },
  {
    "slug": "the-end-of-awkward-ai-transcriptions",
    "title": "The End of Awkward AI Transcriptions",
    "url": "https://ai.engineer/talks/the-end-of-awkward-ai-transcriptions",
    "videoId": "aDj9sY2RoG8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 984000,
    "speakers": [
      {
        "slug": "travis-bartley",
        "name": "Travis Bartley",
        "organization": "NVIDIA"
      },
      {
        "slug": "myungjong-kim",
        "name": "Myungjong Kim",
        "organization": "NVIDIA"
      },
      {
        "slug": "byungjoong",
        "name": "Byungjoong",
        "organization": "NVIDIA"
      },
      {
        "slug": "jaehan",
        "name": "Jaehan",
        "organization": "NVIDIA"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "NVIDIA presenters explain how Riva speech-AI systems combine FastConformer encoders with CTC, RNN-T, and TDT decoding across Parakeet and Canary model families. They describe Sortformer-based diarization and speaker-kernel integration for target-speaker and multitalker transcription, alongside word boosting, text normalization, voice activity detection, and GPU-efficient speech-data loading."
  },
  {
    "slug": "the-engineer-of-the-future-is-the-person-who-is-able-to-choose-what-is-worth-doing-addy-osmani",
    "title": "The engineer of the future is the person who is able to choose what is worth doing — Addy Osmani",
    "url": "https://ai.engineer/talks/the-engineer-of-the-future-is-the-person-who-is-able-to-choose-what-is-worth-doing-addy-osmani",
    "videoId": "n97BCfyFIvw",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1106000,
    "speakers": [
      {
        "slug": "addy-osmani",
        "name": "Addy Osmani"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Addy Osmani argues that as AI agents automate more software production, engineers create value through judgment, accountable production decisions, and ownership of evidence. He explains agent harnesses and software factories, cites Sonar research on AI-code verification and cleaner repositories, and warns against cognitive debt, cognitive surrender, and orchestration overhead. His central boundary is that agents can produce diffs, tests, logs, and traces, but humans remain responsible for understanding systems and deciding what should ship."
  },
  {
    "slug": "the-era-of-compound-engineering",
    "title": "The Era of Compound Engineering",
    "url": "https://ai.engineer/talks/the-era-of-compound-engineering",
    "videoId": "_ehJyfHg1Vk",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1238000,
    "speakers": [
      {
        "slug": "kieran-klaassen",
        "name": "Kieran Klaassen",
        "organization": "Every/Cora"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Every's Kieran Klaassen explains how building Cora, an agent-native AI email client, led him to compound engineering: capturing repeated corrections, product judgment, and implementation knowledge so each completed task makes future work easier. He describes a solo-engineer stack using Rails, Ruby, and React; human-led problem framing and final judgment around autonomous AI execution; and an open-source Compound Engineering plugin spanning Claude Code, Codex, and Cursor, including its CE Ideate workflow."
  },
  {
    "slug": "the-era-of-unbounded-products-designing-for-multimodal-i-o",
    "title": "The era of unbounded products: Designing for Multimodal I/O",
    "url": "https://ai.engineer/talks/the-era-of-unbounded-products-designing-for-multimodal-i-o",
    "videoId": "5nOLb27hQ5w",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1232000,
    "speakers": [
      {
        "slug": "ben-hylak",
        "name": "Ben Hylak",
        "organization": "Dawn Analytics"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Dawn Analytics founder Ben Hylak argues that multimodal AI and spatial computing create unbounded products whose possible inputs, contexts, and behaviors exceed conventional screen-based software. Drawing on Apple Vision Pro and visionOS, he recommends imposing structure through clear priorities, familiar interfaces, examples, and presets. He closes by discussing sparse autoencoders, Golden Gate Claude, controllable image manipulation, and evaluating whether AI experiences meet individual user needs."
  },
  {
    "slug": "the-eyes-are-the-context-window-to-the-soul-how-windsurf-gets-to-know-you",
    "title": "The Eyes Are The (Context) Window to The Soul: How Windsurf Gets to Know You",
    "url": "https://ai.engineer/talks/the-eyes-are-the-context-window-to-the-soul-how-windsurf-gets-to-know-you",
    "videoId": "jUv5WSPo9fk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 817000,
    "speakers": [
      {
        "slug": "sam-fertig",
        "name": "Sam Fertig",
        "organization": "Windsurf"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Windsurf deployed engineer Sam Fertig explains why effective AI coding assistants need contextual understanding of developer behavior, terminal activity, and codebases rather than simply generating code or enlarging LLM context windows. He introduces Windsurf Editor, IDE plugins, and the Cascade agent, discusses the company's GPU-optimization roots as ExaFunction, and closes with claims about stateless processing and customer-data protections."
  },
  {
    "slug": "the-factory-that-dreams-39-ai-agents-no-framework",
    "title": "The Factory That Dreams: 39 AI Agents, No Framework",
    "url": "https://ai.engineer/talks/the-factory-that-dreams-39-ai-agents-no-framework",
    "videoId": "jtzh-GBXBWc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 598000,
    "speakers": [
      {
        "slug": "rushabh-doshi",
        "name": "Rushabh Doshi",
        "organization": "Machinecraft"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Rushabh Doshi explains how Machinecraft transformed closely held manufacturing knowledge into Ira, an organizational AI system composed of bounded specialist agents coordinated by Athena. The system supports sales and customer qualification, combines vector, graph, and CRM data with multiple model providers, retains human corrections, and consolidates useful knowledge during nightly cycles. He concludes with a values-guided, forkable BrainOS architecture that other businesses can populate with their own operational knowledge."
  },
  {
    "slug": "the-friction-is-your-judgment",
    "title": "The Friction Is Your Judgment",
    "url": "https://ai.engineer/talks/the-friction-is-your-judgment",
    "videoId": "_Zcw_sVF6hU",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1118000,
    "speakers": [
      {
        "slug": "armin-ronacher",
        "name": "Armin Ronacher",
        "organization": "Earendil"
      },
      {
        "slug": "cristina-poncela-cubeiro",
        "name": "Cristina Poncela Cubeiro",
        "organization": "Earendil"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Earendil engineers Armin Ronacher and Cristina Poncela Cubeiro argue that frictionless AI-assisted coding can erode human judgment, encourage skipped reviews, and create security or maintenance risks. They distinguish psychological dependence on agent-generated momentum from architectural problems in complex products, then recommend agent-legible codebases with explicit boundaries, centralized SQL access, shared UI primitives, familiar patterns, and TypeScript’s erasableSyntaxOnly mode."
  },
  {
    "slug": "the-future-is-domain-specific-agents",
    "title": "The Future Is Domain-Specific Agents",
    "url": "https://ai.engineer/talks/the-future-is-domain-specific-agents",
    "videoId": "spNAUEgq_A8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1838000,
    "speakers": [
      {
        "slug": "justin-schroeder",
        "name": "Justin Schroeder",
        "organization": "StandardAgents"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "StandardAgents co-founder Justin Schroeder argues that businesses should compose narrowly scoped, domain-specific agents instead of overloading a single general-purpose agent with tools, MCP servers, and skills. He examines enterprise integration challenges, durable execution, accumulating context from workplace applications, inference-token economics, and specialized legal or GDPR-compliance agents, while introducing his open-source dmux and ArrowJS projects."
  },
  {
    "slug": "the-future-of-evals-ankur-goyal-braintrust",
    "title": "The Future of Evals - Ankur Goyal, Braintrust",
    "url": "https://ai.engineer/talks/the-future-of-evals-ankur-goyal-braintrust",
    "videoId": "MC55hdWLq4o",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 314000,
    "speakers": [
      {
        "slug": "ankur-goyal",
        "name": "Ankur Goyal",
        "organization": "Braintrust"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Ankur Goyal describes the growing volume and manual burden of AI evaluations, reporting that typical Braintrust organizations run nearly 13 evaluations daily while some customers exceed 3,000. He introduces Loop, a Braintrust agent that improves prompts, evaluation datasets, and scorers together, displays proposed edits directly in the interface, and supports selectable models including Claude 4, OpenAI models, and Gemini."
  },
  {
    "slug": "the-future-of-evals-from-llm-as-a-judge-to-agent-as-a-judge-aparna-dhinakaran-arize-ai",
    "title": "The Future of Evals: From LLM as a Judge to Agent as a Judge — Aparna Dhinakaran, Arize AI",
    "url": "https://ai.engineer/talks/the-future-of-evals-from-llm-as-a-judge-to-agent-as-a-judge-aparna-dhinakaran-arize-ai",
    "videoId": "q2JrUKBMf0w",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 366000,
    "speakers": [
      {
        "slug": "aparna-dhinakaran",
        "name": "Aparna Dhinakaran",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Arize AI co-founder Aparna Dhinakaran argues that evaluation methods must evolve alongside agents that use tools, reason across long trajectories, and generate dynamic interfaces. Drawing on Arize's Alyx agent and production-trace monitoring, she explains why deterministic checks and fixed-rubric LLM judges remain useful but cannot capture every emerging failure. She introduces Agent-as-a-Judge and Signal as complementary agentic approaches that inspect traces, discover unexpected failure patterns, and help propose fixes through pull requests."
  },
  {
    "slug": "the-future-of-qwen-a-generalist-agent-model",
    "title": "The Future of Qwen: A Generalist Agent Model",
    "url": "https://ai.engineer/talks/the-future-of-qwen-a-generalist-agent-model",
    "videoId": "b0xlsQ_6wUQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1514000,
    "speakers": [
      {
        "slug": "junyang-lin",
        "name": "Junyang Lin",
        "organization": "Alibaba Group / Qwen"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Junyang Lin of Alibaba's Qwen team presents the roadmap toward generalist models and agents, focusing on Qwen3's dense and mixture-of-experts architectures, reinforcement-learning-driven reasoning, hybrid thinking and non-thinking modes, configurable token budgets, multilingual support, and MCP-enabled agent capabilities. He also highlights Qwen Chat's multimodal features, open model checkpoints, and efficient deployment."
  },
  {
    "slug": "the-future-of-work",
    "title": "The Future of Work",
    "url": "https://ai.engineer/talks/the-future-of-work",
    "videoId": "ZYoZSU58m_Y",
    "event": "AI Engineer Summit 2023",
    "durationMs": 773000,
    "speakers": [
      {
        "slug": "toran-bruce-richards",
        "name": "Toran Bruce Richards",
        "organization": "AutoGPT"
      },
      {
        "slug": "silen-naihin",
        "name": "Silen Naihin",
        "organization": "AutoGPT"
      },
      {
        "slug": "poots",
        "name": "Poots",
        "organization": "AutoGPT"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "AutoGPT creator Toran Bruce Richards, founding AI engineer Silen Naihin, and another contributor discuss autonomous agents as a way to reduce repetitive work and expand human creativity. They describe AutoGPT's open-source community, GitHub growth, Agent Protocol-compatible developer interfaces and benchmarks, noisy but improving agent evaluations in continuous integration, prompt-injection risks highlighted by OWASP, and their goal of building an open-source generalist agent."
  },
  {
    "slug": "the-genai-maturity-curve-or-you-probably-don-t-need-fine-tuning",
    "title": "The GenAI Maturity Curve (or: You Probably Don’t Need Fine-Tuning)",
    "url": "https://ai.engineer/talks/the-genai-maturity-curve-or-you-probably-don-t-need-fine-tuning",
    "videoId": "-hYqt8M9u_M",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1083000,
    "speakers": [
      {
        "slug": "kyle-corbitt",
        "name": "Kyle Corbitt",
        "organization": "OpenPipe"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "OpenPipe founder Kyle Corbitt explains when a production GenAI application should progress from prompting a general-purpose model to fine-tuning a task-specific model. He compares output quality, cost, latency, and consistency; discusses multi-stage inference and fine-tuned Llama 3 models; and emphasizes collecting real production prompts, curating representative training data, evaluating results, and deploying only when specialization addresses a demonstrated need."
  },
  {
    "slug": "the-geopolitics-of-ai-infrastructure",
    "title": "The Geopolitics of AI Infrastructure",
    "url": "https://ai.engineer/talks/the-geopolitics-of-ai-infrastructure",
    "videoId": "Zz4QjZsYWK0",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1109000,
    "speakers": [
      {
        "slug": "dylan-patel",
        "name": "Dylan Patel",
        "organization": "SemiAnalysis"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "SemiAnalysis founder Dylan Patel examines AI infrastructure as a geopolitical contest spanning China, the United States, and the Middle East. He compares Huawei Ascend and CloudMatrix architectures with NVIDIA Blackwell and NVLink systems, discusses semiconductor manufacturing, high-bandwidth memory, and export-control exposure, then considers Middle Eastern compute investment, Stargate-related projects, converging training and inference infrastructure, and power constraints. The closing audience question addresses SMIC's reliance on foreign chipmaking equipment."
  },
  {
    "slug": "the-golden-age-of-ai-engineering",
    "title": "The Golden Age of AI Engineering",
    "url": "https://ai.engineer/talks/the-golden-age-of-ai-engineering",
    "videoId": "pMggiOb18tc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1513000,
    "speakers": [
      {
        "slug": "alexander-embiricos",
        "name": "Alexander Embiricos",
        "organization": "OpenAI"
      },
      {
        "slug": "romain-huet",
        "name": "Romain Huet",
        "organization": "OpenAI"
      },
      {
        "slug": "peter-steinberger",
        "name": "Peter Steinberger",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "OpenAI presenters Alexander Embiricos and Romain Huet argue that increasingly capable AI systems expand the importance of engineering rather than eliminate it. They contrast earlier DevDay demonstrations with newer autonomous build-and-test capabilities, explain Codex’s chat-first collaborative design and its shared Responses API and context-compaction primitives, and highlight third-party applications built on the Codex app-server. Guest Peter Steinberger closes by emphasizing stronger agent harnesses, orchestration, and engineering loops over manually juggling terminal sessions."
  },
  {
    "slug": "the-great-loops-debate-dex-horthy-geoff-huntley-ian-livingstone-greg-pstrucha-moderated-by-allie",
    "title": "The Great Loops Debate — Dex Horthy, Geoff Huntley, Ian Livingstone, Greg Pstrucha, moderated by Allie Howe",
    "url": "https://ai.engineer/talks/the-great-loops-debate-dex-horthy-geoff-huntley-ian-livingstone-greg-pstrucha-moderated-by-allie",
    "videoId": "c35YoMdnI78",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 3616000,
    "speakers": [
      {
        "slug": "allie-howe",
        "name": "Allie Howe",
        "organization": "Keycard"
      },
      {
        "slug": "dex-horthy",
        "name": "Dex Horthy",
        "organization": "HumanLayer"
      },
      {
        "slug": "geoffrey-huntley",
        "name": "Geoffrey Huntley"
      },
      {
        "slug": "ian-livingstone",
        "name": "Ian Livingstone",
        "organization": "Keycard"
      },
      {
        "slug": "greg-pstrucha",
        "name": "Greg Pstrucha",
        "organization": "Sentry"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Moderator Allie Howe leads an Oxford-style debate between Ian Livingstone and Geoffrey Huntley, who argue that disciplined coding-agent loops already deliver practical autonomy, and Dex Horthy and Greg Pstrucha, who challenge the gap between loop hype and reliable engineering outcomes. The discussion examines Ralph loops, context windows, verification and code review, CI/CD, pre-commit back pressure, persistent shared state, multi-agent coordination, and the prospects for software factories."
  },
  {
    "slug": "the-hidden-costs-of-building-your-own-rag-stack-ofer-mendelevitch-vectara",
    "title": "The Hidden Costs of Building Your Own RAG Stack — Ofer Mendelevitch, Vectara",
    "url": "https://ai.engineer/talks/the-hidden-costs-of-building-your-own-rag-stack-ofer-mendelevitch-vectara",
    "videoId": "62U6FLUCPWs",
    "event": "AI Engineer Summit 2025",
    "durationMs": 914000,
    "speakers": [
      {
        "slug": "ofer-mendelevitch",
        "name": "Ofer Mendelevitch",
        "organization": "Vectara"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Vectara developer-relations leader Ofer Mendelevitch explains how retrieval-augmented generation grounds LLM responses in enterprise data, then examines the hidden complexity of building an enterprise-grade RAG stack. He contrasts self-managed ingestion, parsing, chunking, retrieval, hallucination detection, provenance, latency optimization, scaling, access controls, and operational staffing with Vectara's managed, API-accessible RAG platform and its SaaS, VPC, and on-premises deployment options."
  },
  {
    "slug": "the-hidden-life-of-embeddings",
    "title": "The Hidden Life of Embeddings",
    "url": "https://ai.engineer/talks/the-hidden-life-of-embeddings",
    "videoId": "YvobVu1l7GI",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1095000,
    "speakers": [
      {
        "slug": "linus-lee",
        "name": "Linus Lee",
        "organization": "Notion"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Notion research engineer Linus Lee demonstrates how embeddings can power exploratory interfaces beyond vector search, including semantic direction editing, sentiment manipulation, vector splicing, and text reconstruction. He adapts a T5-based denoising autoencoder to recover text from OpenAI text-embedding-ada-002 vectors, then uses CLIP and Kakao Brain’s Karlo unCLIP diffusion model for image interpolation and multimodal vector arithmetic."
  },
  {
    "slug": "the-hierarchy-of-needs-for-training-dataset-development",
    "title": "The Hierarchy of Needs for Training Dataset Development",
    "url": "https://ai.engineer/talks/the-hierarchy-of-needs-for-training-dataset-development",
    "videoId": "i2vBaFzCEJw",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 992000,
    "speakers": [
      {
        "slug": "chang-she",
        "name": "Chang She",
        "organization": "LanceDB"
      },
      {
        "slug": "noah-shpak",
        "name": "Noah Shpak",
        "organization": "Character.ai"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Chang She of LanceDB and Noah Shpak of Character.ai explain how foundation-model training depends on continuously inspecting, constructing, and refining large multimodal datasets. They describe Character.ai’s research platform, including Spark, Trino, GPU-backed synthetic-data workflows, and dataset materialization, before showing how Lance supports inexpensive row-reference shuffling, fast random access, scans, large-binary streaming, and zero-copy schema evolution."
  },
  {
    "slug": "the-infrastructure-for-the-singularity",
    "title": "The infrastructure for the singularity",
    "url": "https://ai.engineer/talks/the-infrastructure-for-the-singularity",
    "videoId": "2goSS66XRBk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1171000,
    "speakers": [
      {
        "slug": "jesse-han",
        "name": "Jesse Han",
        "organization": "Morph Labs"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Jesse Han presents Morph Labs’ vision for agent-native cloud infrastructure, describing Infinibranch as a virtualization, storage, and networking stack that enables reversible execution, snapshots, and parallel reasoning. He announces Morph Liquid Metal, highlighting faster branching, container support, elastic autoscaling, and planned GPU support, then connects agent swarms and verified reasoning to Morph Cloud’s broader superintelligence roadmap."
  },
  {
    "slug": "the-intelligent-interface",
    "title": "The Intelligent Interface",
    "url": "https://ai.engineer/talks/the-intelligent-interface",
    "videoId": "ieWT6X2Yh_g",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1136000,
    "speakers": [
      {
        "slug": "samantha-whitmore",
        "name": "Samantha Whitmore",
        "organization": "New Computer"
      },
      {
        "slug": "jason-yuan",
        "name": "Jason Yuan",
        "organization": "New Computer"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "New Computer co-founders Samantha Whitmore and Jason Yuan rethink AI interfaces from first principles, moving beyond text-only conversation toward systems that sense context, reason, and respond through appropriate modalities. A live demonstration combines camera-based pose detection, LLM-selected input and output, and continuously updated world state; subsequent speculative concepts explore headset hand and eye tracking, social interfaces, and a Midjourney-like generative iPad canvas grounded in physical surroundings."
  },
  {
    "slug": "the-knowledge-graph-mullet-trimming-graphrag-complexity",
    "title": "The Knowledge Graph Mullet: Trimming GraphRAG Complexity",
    "url": "https://ai.engineer/talks/the-knowledge-graph-mullet-trimming-graphrag-complexity",
    "videoId": "tYCu_57jzL8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1972000,
    "speakers": [
      {
        "slug": "william-lyon",
        "name": "William Lyon",
        "organization": "Hypermode"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "William Lyon of Hypermode presents a hybrid knowledge-graph architecture that combines property-graph interfaces with RDF triple storage to simplify GraphRAG. Using Dgraph, he explains RDF modeling, posting-list optimizations, DQL and GraphQL-inspired queries, Ratel, and vector search, then demonstrates Model Context Protocol integration, graph-powered recommendations, and long-running Modus agents that retrieve GitHub content and generate code-aware outputs."
  },
  {
    "slug": "the-last-human-code-review-building-trust-in-ai-generated-code",
    "title": "The Last Human Code Review: Building Trust in AI-Generated Code",
    "url": "https://ai.engineer/talks/the-last-human-code-review-building-trust-in-ai-generated-code",
    "videoId": "s-aixZYJG4c",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1134000,
    "speakers": [
      {
        "slug": "itamar-friedman",
        "name": "Itamar Friedman",
        "organization": "Qodo"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Qodo co-founder and CEO Itamar Friedman argues that AI-assisted software development shifts the bottleneck from generating code to establishing trust in code review. He describes consolidating fragmented team standards, architectural knowledge, agent instructions, and historical context into governance systems usable by both developers and AI agents. As review quality improves and human intervention declines, organizations can gradually automate pull-request approval and blocking, ultimately evaluating concurrent changes and service contracts across a software graph rather than isolated diffs."
  },
  {
    "slug": "the-llm-triangle-engineering-principles-for-robust-ai-applications",
    "title": "The LLM Triangle: Engineering Principles for Robust AI Applications",
    "url": "https://ai.engineer/talks/the-llm-triangle-engineering-principles-for-robust-ai-applications",
    "videoId": "ro5HkZvzfiQ",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1579000,
    "speakers": [
      {
        "slug": "almog-baku",
        "name": "Almog Baku",
        "organization": "GenAI Israel"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "In this prerecorded AI Engineer Summit online talk, GenAI Israel founder Almog Baku presents the LLM Triangle, a production-engineering framework combining foundation models, engineering techniques, and data under explicit standard operating procedures. He explains why proofs of concept fail with real users, compares prompting and software techniques including chain of thought, few-shot examples, ReAct, and structured output, and illustrates LLM-native workflows and tool-using agents through a Wikipedia-writing example. He emphasizes data quality as central to reliable application performance."
  },
  {
    "slug": "the-log-is-the-agent",
    "title": "The Log Is The Agent",
    "url": "https://ai.engineer/talks/the-log-is-the-agent",
    "videoId": "UPwGaM2MKHY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 911000,
    "speakers": [
      {
        "slug": "ishaan-sehgal",
        "name": "Ishaan Sehgal",
        "organization": "Omnara"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "Omnara CEO Ishaan Sehgal argues that an AI agent's durable identity resides in its session log rather than its model, runtime, tools, or machine. Using a Skyrim save-file analogy, he describes model context and interfaces as projections of persistent agent history, enabling crash recovery, preserved permission prompts, branching, and portability across Claude, GPT, and Qwen. He critiques fragile local JSONL and SQLite persistence and warns that providers controlling agent logs can create the deepest form of infrastructure lock-in."
  },
  {
    "slug": "the-making-of-devin",
    "title": "The Making of Devin",
    "url": "https://ai.engineer/talks/the-making-of-devin",
    "videoId": "T7NWjoD_OuY",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1204000,
    "speakers": [
      {
        "slug": "scott-wu",
        "name": "Scott Wu",
        "organization": "Cognition (Devin)"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Cognition co-founder and CEO Scott Wu demonstrates Devin building and deploying a mobile-friendly React name game from a TSV of conference speakers. He explains Devin's planning, repository-aware Git and pull-request workflows, internal product development, context management, Slack and GitHub integrations, verification tooling, and asynchronous engineering collaboration, before taking questions about access and implications for junior developers."
  },
  {
    "slug": "the-many-ends-of-programming",
    "title": "The Many Ends of Programming",
    "url": "https://ai.engineer/talks/the-many-ends-of-programming",
    "videoId": "5s6Q-y42ZZA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1632000,
    "speakers": [
      {
        "slug": "ray-myers",
        "name": "Ray Myers",
        "organization": "All Hands AI"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Ray Myers, identifying himself as chief architect at All Hands AI, examines competing futures for software development under AI and argues for empathy rather than deterministic claims about programmers becoming obsolete. He contrasts legacy-code maintenance with autocomplete tools such as Cursor and GitHub Copilot, demonstrates Haskell-assisted code generation and an OpenHands agent producing a broad pull request, and considers type theory, domain-specific languages, predictable abstractions, software quality, and humane collaboration."
  },
  {
    "slug": "the-maturity-phases-of-running-evals",
    "title": "The maturity phases of running evals",
    "url": "https://ai.engineer/talks/the-maturity-phases-of-running-evals",
    "videoId": "FB-MLPhL9Ms",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1114000,
    "speakers": [
      {
        "slug": "phil-hetzel",
        "name": "Phil Hetzel",
        "organization": "Braintrust"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Braintrust solutions-engineering leader Phil Hetzel describes how agent-evaluation practices mature from focused human assessments into scalable, production-informed quality systems. He recommends documenting subject-matter experts’ judgments, converting those rationales into scoring functions and LLM-as-a-judge evaluations, validating automated judges, and expanding assessment from individual outputs to complete agent traces. He concludes by highlighting production-scale topic modeling for discovering failure modes and discusses deterministic evaluation during audience questions."
  },
  {
    "slug": "the-messy-reality-of-scale-synthetic-data-and-pre-training",
    "title": "The Messy Reality of Scale: Synthetic Data and Pre-Training",
    "url": "https://ai.engineer/talks/the-messy-reality-of-scale-synthetic-data-and-pre-training",
    "videoId": "KhYifX22yhE",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1051000,
    "speakers": [
      {
        "slug": "marah-abdin",
        "name": "Marah Abdin",
        "organization": "Poolside"
      },
      {
        "slug": "robert-mchardy",
        "name": "Robert McHardy",
        "organization": "Poolside"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Marah Abdin and Robert McHardy describe how poolside combines synthetic-data generation with large-scale pre-training for its Laguna coding models. Abdin explains rephrasing, code and STEM pipelines, multi-agent orchestration, validation, and a 13% synthetic-data share in Laguna XS.2 pre-training. McHardy details failures involving defective GPUs, silent data corruption, exploding gradients, and insufficient BF16 precision during tensor-parallel accumulation, before previewing the larger open-weight Laguna S model."
  },
  {
    "slug": "the-miranda-hypothesis-how-hamilton-the-musical-poisoned-your-persona-evals",
    "title": "The Miranda Hypothesis: How Hamilton (the Musical) Poisoned Your Persona Evals",
    "url": "https://ai.engineer/talks/the-miranda-hypothesis-how-hamilton-the-musical-poisoned-your-persona-evals",
    "videoId": "IJXjTLPzvAU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 3497000,
    "speakers": [
      {
        "slug": "jacob-e-thomas",
        "name": "Jacob E. Thomas",
        "organization": "Results Generation"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Jacob E. Thomas examines how role-playing language agents reproduce culturally popular composites of historical figures instead of reasoning from documents available at a specific historical moment. Using Alexander Hamilton, Abraham Lincoln and the InCharacter benchmark, he argues that personality fidelity, human-preference optimization and fine-tuning can conceal anachronistic reasoning. He outlines a preregistered, temporally anchored evaluation using historical source material, diagnostic questions, documentary consistency and contextual plausibility."
  },
  {
    "slug": "the-missing-layer-after-launch",
    "title": "The Missing Layer After Launch",
    "url": "https://ai.engineer/talks/the-missing-layer-after-launch",
    "videoId": "kZsf_Sfm7RU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1173000,
    "speakers": [
      {
        "slug": "raphael-kalandadze",
        "name": "Raphael Kalandadze",
        "organization": "Wandero AI"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Raphael Kalandadze explains why launching an AI agent begins a distinct operational challenge: monitoring real customer conversations, detecting failures that conventional tests and logs can miss, and managing non-deterministic behavior. He describes an agent-operated loop that analyzes traces and logs, diagnoses root causes, drafts pull requests, subjects fixes to independent review agents, and escalates decisions to humans. Operational dashboards track analyzed sessions, costs, success rates, trends, and emerging patterns."
  },
  {
    "slug": "the-missing-layer-design-taste-in-ai-agents",
    "title": "The Missing Layer: Design Taste in AI Agents",
    "url": "https://ai.engineer/talks/the-missing-layer-design-taste-in-ai-agents",
    "videoId": "7GMKdpLsxwU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 850000,
    "speakers": [
      {
        "slug": "hassan-el-mghari",
        "name": "Hassan El Mghari",
        "organization": "Together AI"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Together AI developer-experience lead Hassan El Mghari argues that deliberate design and UX distinguish useful AI-built applications from generic agent-generated interfaces. He demonstrates applications built with open-source models, identifies recurring visual anti-patterns, and introduces Hallmark, an open-source design skill that combines explicit anti-slop rules with strong visual references. He also discusses GLM 5.2, persistent preferences in AGENTS.md, and treating generated interfaces as a starting point for iterative refinement."
  },
  {
    "slug": "the-missing-primitive-for-agent-swarms",
    "title": "The Missing Primitive for Agent Swarms",
    "url": "https://ai.engineer/talks/the-missing-primitive-for-agent-swarms",
    "videoId": "5Sui_OnSRlY",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1117000,
    "speakers": [
      {
        "slug": "lou-bichard",
        "name": "Lou Bichard",
        "organization": "Ona"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Ona field CTO Lou Bichard examines software factories built from autonomous coding-agent swarms and fleets, contrasting Stripe's Minions and Ramp's Inspect with Ona's development-environment platform. He discusses repository-encoded harness engineering, event-driven automation, scalable remediation, context limitations, and runtime isolation with worktrees, containers, and VMs, then demonstrates Ona and explores the unresolved coordination layer and agent-protocol questions."
  },
  {
    "slug": "the-model-isn-t-wrong-you-re-just-bad-at-prompting",
    "title": "The Model Isn’t Wrong—You’re Just Bad at Prompting",
    "url": "https://ai.engineer/talks/the-model-isn-t-wrong-you-re-just-bad-at-prompting",
    "videoId": "Hp4MzVTXcKw",
    "event": "AI Engineer Summit 2025",
    "durationMs": 534000,
    "speakers": [
      {
        "slug": "dan-cleary",
        "name": "Dan Cleary",
        "organization": "PromptHub"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "PromptHub co-founder Dan Cleary explains why prompt engineering remains a practical first step before introducing more complex LLM architectures. He covers chain-of-thought and few-shot prompting, automatically generated reasoning chains, and model-provider-specific meta-prompting. Using examples involving DeepSeek-R1, OpenAI o1, and MedPrompt, he discusses why reasoning models may benefit from extended reasoning but can perform worse when supplied unnecessary few-shot examples or excessive context."
  },
  {
    "slug": "the-multimodal-future-of-education",
    "title": "The Multimodal Future of Education",
    "url": "https://ai.engineer/talks/the-multimodal-future-of-education",
    "videoId": "qpmZID27t98",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1205000,
    "speakers": [
      {
        "slug": "stefania-druga",
        "name": "Stefania Druga",
        "organization": "Google"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Google researcher Stefania Druga explores how multimodal and generative AI can support education while preserving learners’ agency and strengthening critical AI literacy. Drawing on research into children’s perceptions of voice assistants, she introduces Cognimates, a Scratch-based platform where young people program robots and train models, and describes how hands-on experimentation and family collaboration can foster a more critical understanding of AI."
  },
  {
    "slug": "the-new-application-layer",
    "title": "The New Application Layer",
    "url": "https://ai.engineer/talks/the-new-application-layer",
    "videoId": "XKup1pj-34M",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1132000,
    "speakers": [
      {
        "slug": "malte-ubl",
        "name": "Malte Ubl",
        "organization": "Vercel"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Vercel CTO Malte Ubl argues that AI engineering is succeeding web development and that agents make previously uneconomical software practical. He introduces Chat SDK and just-bash, discusses implications for software demand and SaaS, reports that AI agents generated more than 60% of recent vercel.com page views, and describes the resulting shift toward APIs, CLIs, agent-oriented infrastructure, and application-layer innovation exemplified by OpenClaw."
  },
  {
    "slug": "the-new-lean-startup",
    "title": "The New Lean Startup",
    "url": "https://ai.engineer/talks/the-new-lean-startup",
    "videoId": "pQz-PgA1eJw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 806000,
    "speakers": [
      {
        "slug": "sid-bendre",
        "name": "Sid Bendre",
        "organization": "Oleve"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Oleve co-founder Sid Bendre explains how a four-person team built a profitable consumer-software portfolio, describing Quizard AI and Unstuck AI, early product distribution, generalist hiring, KPI ownership, and product-engineer accountability. He details using LaunchDarkly to route LLM traffic around provider rate limits, recover from third-party failures, and run interface or paywall experiments without deployments, then outlines AI-augmented operations and briefly previews a separate TRELLIS presentation."
  },
  {
    "slug": "the-new-primitives-building-ai-native-software",
    "title": "The New Primitives: Building AI-Native Software",
    "url": "https://ai.engineer/talks/the-new-primitives-building-ai-native-software",
    "videoId": "LZuWZRze3MU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1274000,
    "speakers": [
      {
        "slug": "kwindla-kramer",
        "name": "Kwindla Kramer",
        "organization": "Daily"
      },
      {
        "slug": "kwindla-hultman-kramer",
        "name": "Kwindla Hultman Kramer",
        "organization": "Daily"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Daily's Kwindla Hultman Kramer introduces Pipecat and argues that today's AI agents are an intermediate step toward richer AI-native software. Drawing on Vannevar Bush's 1945 essay, computing history, conversational science-fiction interfaces, and a modern Knowledge Navigator-style demonstration, he examines how models, data, tools, context engineering, and multimodal voice interfaces can form the next generation of software primitives."
  },
  {
    "slug": "the-next-game-engine-won-t-have-a-manual",
    "title": "The Next Game Engine Won't Have a Manual",
    "url": "https://ai.engineer/talks/the-next-game-engine-won-t-have-a-manual",
    "videoId": "VBCDhRrvlYo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1173000,
    "speakers": [
      {
        "slug": "arturo-nunez",
        "name": "Arturo Nunez",
        "organization": "Nereu"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "Nereu founder Arturo Nunez demonstrates an AI-native game-creation workflow in which the Bibi assistant translates natural-language requests into assets, movement, animation, and gameplay behavior. He argues that conventional engines burden creators with repetitive components and boilerplate, and presents an asset tag system that lets engine systems infer behavior from tagged objects. The browser-based JavaScript implementation also uses vision models to describe and tag thousands of 3D assets, aiming to make game development more accessible and enjoyable."
  },
  {
    "slug": "the-next-medium-why-real-time-interactive-video-changes-everything-ahmed-ahres-reactor",
    "title": "The Next Medium: Why Real-Time Interactive Video Changes Everything — Ahmed Ahres, Reactor",
    "url": "https://ai.engineer/talks/the-next-medium-why-real-time-interactive-video-changes-everything-ahmed-ahres-reactor",
    "videoId": "5dCAmSDOAjI",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1050016,
    "speakers": [
      {
        "slug": "ahmed-ahres",
        "name": "Ahmed Ahres",
        "organization": "Reactor"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "Reactor's Ahmed Ahres argues that world models are best understood as real-time, interactive video rather than static or batch-generated media. He contrasts continuous, controllable generation with conventional video models; discusses applications in advertising, immersive education, audience-directed livestreaming, and video editing; and introduces Reactor's developer platform and Helios model. Audience questions address interactive generation frame rates and ongoing research."
  },
  {
    "slug": "the-next-unicorns-7-top-ai-startups-from-the-hf0-residency",
    "title": "The Next Unicorns: 7 Top AI startups from the HF0 Residency",
    "url": "https://ai.engineer/talks/the-next-unicorns-7-top-ai-startups-from-the-hf0-residency",
    "videoId": "L8-5ezsoI5A",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1336000,
    "speakers": [
      {
        "slug": "diego-rodriguez",
        "name": "Diego Rodriguez",
        "organization": "Krea"
      },
      {
        "slug": "eugene",
        "name": "Eugene",
        "organization": "Featherless.ai"
      },
      {
        "slug": "jonas-bauer",
        "name": "Jonas Bauer",
        "organization": "Upside"
      },
      {
        "slug": "shijia-liao",
        "name": "Shijia Liao",
        "organization": "OpenAudio / Fish Audio"
      },
      {
        "slug": "david-vorick",
        "name": "David Vorick",
        "organization": "Glow"
      },
      {
        "slug": "alex-atallah",
        "name": "Alex Atallah",
        "organization": "OpenRouter"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "An HF0 Residency startup showcase features rapid pitches covering Krea's AI creative suite, OpenHome's programmable smart speakers, Featherless AI, Upside's structured enterprise-data platform, OpenAudio/Fish Audio's instruction-controlled S1 voice model, Glow's token-incentivized solar deployment, and OpenRouter's model-routing roadmap. The description also lists Coframe, while David Vorick's Glow presentation appears in the transcript but is absent from the supplied timestamp list; presenter identities and the proposed conference-session match therefore require review."
  },
  {
    "slug": "the-pipeline-is-dead",
    "title": "The Pipeline Is Dead",
    "url": "https://ai.engineer/talks/the-pipeline-is-dead",
    "videoId": "bRnoEpoK5m4",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1190000,
    "speakers": [
      {
        "slug": "iris-ten-teije",
        "name": "Iris ten Teije",
        "organization": "Sky Valley Ambient Computing"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Iris ten Teije argues that traditional CI and distribution pipelines were designed around shipping one frozen software artifact to every user, an assumption challenged by inexpensive AI-assisted changes at runtime. She introduces Differ, Sky Valley Ambient Computing's adaptive-software product, and describes a shared software stem with bounded, isolated, reversible per-user divergences. Examples include enterprise customization and role-specific CRM experiences, alongside the unresolved infrastructure challenges of testing, correctness, provenance, validation, and measuring whether personalized changes improve meaningful outcomes."
  },
  {
    "slug": "the-price-of-intelligence-ai-agent-pricing-in-2025",
    "title": "The Price of Intelligence - AI Agent Pricing in 2025",
    "url": "https://ai.engineer/talks/the-price-of-intelligence-ai-agent-pricing-in-2025",
    "videoId": "In7K-4JZKR4",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1238000,
    "speakers": [
      {
        "slug": "kshitij-grover",
        "name": "Kshitij Grover",
        "organization": "Orb"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Orb co-founder and CTO Kshitij Grover explains how AI-agent companies can align pricing with customer value, buyer behavior, operating costs, and changing product usage. Examples include Intercom Fin's outcome-based pricing, Unify's credits, Cursor's usage limits and premium models, Chargeflow's recovery-based fees, and enterprise-oriented sales approaches from HeavyAI and ServiceNow. He emphasizes audience segmentation, margins, pricing flexibility, and billing infrastructure that supports frequent price changes, versioning, and migrations."
  },
  {
    "slug": "the-production-ai-playbook-deploying-agents-at-enterprise-scale",
    "title": "The Production AI Playbook: Deploying Agents at Enterprise Scale",
    "url": "https://ai.engineer/talks/the-production-ai-playbook-deploying-agents-at-enterprise-scale",
    "videoId": "ObTPqBGsEbA",
    "event": "AI Engineer Europe 2026",
    "durationMs": 2226000,
    "speakers": [
      {
        "slug": "sandipan-bhaumik",
        "name": "Sandipan Bhaumik",
        "organization": "Databricks"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "finance",
        "name": "Finance"
      }
    ],
    "summary": "Databricks technical lead Sandipan Bhaumik explains how to move enterprise AI agents from demos into production by defining measurable business outcomes, building continuously updated evaluation datasets, instrumenting tracing and observability, establishing governed data foundations, and preparing operational ownership. A retail banking chatbot case study illustrates delaying model selection until evaluation infrastructure exists, targeting measurable query deflection, and monitoring accuracy, response time, and customer satisfaction after launch. The talk highlights MLflow LLM judges, Unity Catalog, Mosaic AI, PII detection, prompt versioning, and related multi-agent orchestration."
  },
  {
    "slug": "the-prompt-is-still-a-punch-card",
    "title": "The Prompt Is Still a Punch Card",
    "url": "https://ai.engineer/talks/the-prompt-is-still-a-punch-card",
    "videoId": "hVJOnuhFmTA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1213000,
    "speakers": [
      {
        "slug": "ted-johnson",
        "name": "Ted Johnson",
        "organization": "JoinIn AI"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "JoinIn AI co-founder Ted Johnson argues that modern prompting retains the batch-processing protocol of punch cards: people must assemble complete requests, submit them, wait, and retry. Tracing inherited interface constraints through keyboard layouts and the distinction between communication channels, expression, and protocol, he contrasts rapidly advancing language-model capabilities with stagnant interaction design. Speech and meeting demonstrations illustrate the need for AI that handles interruptions, multiple speakers, references, ambiguity, timing, and conversational context, allowing interfaces to adapt to people rather than requiring people to adapt to machines."
  },
  {
    "slug": "the-prompt-is-the-platform",
    "title": "The Prompt is the Platform",
    "url": "https://ai.engineer/talks/the-prompt-is-the-platform",
    "videoId": "DqtmZE6Hl0g",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1053000,
    "speakers": [
      {
        "slug": "dominik",
        "name": "Dominik",
        "organization": "Resonate HQ, Inc"
      },
      {
        "slug": "dominik-tornow",
        "name": "Dominik Tornow",
        "organization": "Resonate HQ, Inc"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Resonate founder Dominik Tornow argues that coding agents can replace general-purpose software platforms with bespoke implementations generated from abstract specifications and existing infrastructure. Using Resonate's durable-execution protocol and a NATS-based implementation, he contrasts human-led and agent-driven design, describes target-specific decisions for Postgres and NATS primitives, and explains how deterministic Python simulations make distributed-system behavior repeatable and inspectable. His central thesis is that the specification becomes the product and the prompt becomes the platform."
  },
  {
    "slug": "the-rag-stack-we-landed-on-after-37-fails",
    "title": "The RAG Stack We Landed On After 37 Fails",
    "url": "https://ai.engineer/talks/the-rag-stack-we-landed-on-after-37-fails",
    "videoId": "2CXn-CByNoo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1132000,
    "speakers": [
      {
        "slug": "jonathan-fernandes",
        "name": "Jonathan Fernandes"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Independent AI engineer Jonathan Fernandes distills lessons from 37 failed RAG experiments into a practical stack spanning [REDACTED:username] orchestration, open NVIDIA/BAAI embeddings, Qdrant vector storage, locally served Llama or Qwen models, cross-encoder reranking, Arize Phoenix observability, and Ragas evaluation. Using a London railway knowledge base, he contrasts Google Colab prototyping with Docker-based production deployment and demonstrates how retrieval and reranking improvements produce more relevant answers."
  },
  {
    "slug": "the-rise-of-caas-context-as-a-service-for-agentic-ai",
    "title": "The Rise of CaaS: Context-as-a-Service for Agentic AI",
    "url": "https://ai.engineer/talks/the-rise-of-caas-context-as-a-service-for-agentic-ai",
    "videoId": "Ot4OPrPH4xY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1340000,
    "speakers": [
      {
        "slug": "omer-primor",
        "name": "Omer Primor",
        "organization": "Bright Data"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Bright Data product marketing lead Omer Primor examines Context-as-a-Service as infrastructure for AI agents that require fresh, structured web knowledge. He explains why changing web data and point-in-time search limit longitudinal analysis, surveys providers including Exa, Tavily, Google, Claude, and Amazon Bedrock AgentCore, and presents an explicitly informal comparison of search, context vendors, and a quickly assembled in-house pipeline. His central argument is that repeated retrieval frequency drives rental costs, while directly owned context pipelines can create compounding economic advantages."
  },
  {
    "slug": "the-rise-of-open-models-in-the-enterprise",
    "title": "The Rise of Open Models in the Enterprise",
    "url": "https://ai.engineer/talks/the-rise-of-open-models-in-the-enterprise",
    "videoId": "3WV1vT0B0cg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1010000,
    "speakers": [
      {
        "slug": "amir-haghighat",
        "name": "Amir Haghighat",
        "organization": "Baseten"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Baseten co-founder and CTO Amir Haghighat examines how enterprises move from purchasing vertical AI products and experimenting with hosted proprietary models toward building differentiated AI systems. He discusses dedicated Azure and AWS deployments, provider interoperability, proprietary enterprise data, inference-quality constraints, latency-versus-throughput tradeoffs for voice applications, and the operational expertise required to run open-model infrastructure with tools such as SGLang and vLLM."
  },
  {
    "slug": "the-rise-of-the-agentic-economy-on-the-shoulders-of-mcp",
    "title": "The rise of the agentic economy on the shoulders of MCP",
    "url": "https://ai.engineer/talks/the-rise-of-the-agentic-economy-on-the-shoulders-of-mcp",
    "videoId": "blW-lSd5CYQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1088000,
    "speakers": [
      {
        "slug": "jan-curn",
        "name": "Jan Curn",
        "organization": "Apify"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Apify founder Jan Curn argues that collective intelligence can emerge from networks of autonomous agents connected through Model Context Protocol. He explains how Apify exposes its marketplace of approximately 5,000 Docker-based Actors through dynamic MCP tool discovery, avoiding the context burden of publishing every tool at once. The talk examines agent identity, payment, monetization, and trust challenges, then demonstrates invoking Browserbase's MCP capabilities through an Apify Actor to enable cross-service browser automation."
  },
  {
    "slug": "the-robots-are-coming-for-your-job-and-that-s-okay",
    "title": "The Robots Are Coming for Your Job, and That's Okay",
    "url": "https://ai.engineer/talks/the-robots-are-coming-for-your-job-and-that-s-okay",
    "videoId": "pSqpC7fFLZA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 537000,
    "speakers": [
      {
        "slug": "elmer-thomas",
        "name": "Elmer Thomas",
        "organization": "Twilio"
      },
      {
        "slug": "maria-bermudez",
        "name": "Maria Bermudez",
        "organization": "Twilio"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Twilio developer educators Elmer Thomas and Maria Bermudez explain how their documentation team uses six narrowly scoped AI agents to automate editing, alt text, jargon simplification, SEO metadata, documentation outlines, and Slack triage while retaining human judgment. They describe a Next.js interface, Airtable-managed editorial guidance, o1-based editing, Vale linting, CI/CD checks, and GitHub pull-request review; Bermudez demonstrates live-page editing, change explanations, SEO generation, and text simplification."
  },
  {
    "slug": "the-roi-of-ai-why-you-need-eval-frameworks",
    "title": "The ROI of AI: Why You Need Eval Frameworks",
    "url": "https://ai.engineer/talks/the-roi-of-ai-why-you-need-eval-frameworks",
    "videoId": "0oTlgBfzNJ0",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1528000,
    "speakers": [
      {
        "slug": "beyang-liu",
        "name": "Beyang Liu",
        "organization": "Sourcegraph"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Sourcegraph CTO and co-founder Beyang Liu explains how engineering leaders can evaluate the return on AI coding assistants beyond intuition. Using Cody and enterprise customer examples, he describes the importance of code search and relevant codebase context, discusses rigorous evaluation approaches and engineering KPIs, and examines developer-productivity measurement alongside lessons from The Mythical Man-Month."
  },
  {
    "slug": "the-small-model-infrastructure-nobody-built-so-we-did-filip-makraduli-superlinked",
    "title": "The Small Model Infrastructure Nobody Built (So We Did) — Filip Makraduli, Superlinked",
    "url": "https://ai.engineer/talks/the-small-model-infrastructure-nobody-built-so-we-did-filip-makraduli-superlinked",
    "videoId": "qdh_x-uRs9g",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1110000,
    "speakers": [
      {
        "slug": "filip-makraduli",
        "name": "Filip Makraduli",
        "organization": "Superlinked"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Filip Makraduli explains why production embeddings, retrieval, document processing, and agentic workflows need infrastructure designed specifically for small models. Introducing Superlinked Inference Engine (SIE), he describes loading and hot-swapping multiple models on shared GPUs, improving utilization, and combining model support with production routing, autoscaling, queuing, and GPU provisioning. Examples include Stella, GLiNER, vector-database integrations, context rot, benchmark evaluation, token-padding inefficiencies, and a closing discussion of sinusoidal positional encodings."
  },
  {
    "slug": "the-state-of-ai-code-quality-hype-vs-reality",
    "title": "The State of AI Code Quality: Hype vs. Reality",
    "url": "https://ai.engineer/talks/the-state-of-ai-code-quality-hype-vs-reality",
    "videoId": "rgjF5o2Qjsc",
    "event": "AI Engineer Code 2025",
    "durationMs": 1275000,
    "speakers": [
      {
        "slug": "itamar-friedman",
        "name": "Itamar Friedman",
        "organization": "Qodo"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Qodo co-founder and CEO Itamar Friedman contrasts AI-coding productivity claims with practical software-quality risks. Drawing on Qodo's developer research, he discusses adoption, unreliable coding rules and security reviews, growing pull-request and review burdens, and the need for contextual standards, testing, verification, code ownership, and agentic quality workflows across the software development lifecycle."
  },
  {
    "slug": "the-state-of-ai-powered-search-and-retrieval",
    "title": "The State of AI-Powered Search and Retrieval",
    "url": "https://ai.engineer/talks/the-state-of-ai-powered-search-and-retrieval",
    "videoId": "pIPtpBZ6TKk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 755000,
    "speakers": [
      {
        "slug": "frank-liu",
        "name": "Frank Liu",
        "organization": "MongoDB"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Frank Liu of MongoDB and the former Voyage AI team explains how semantic, embedding-based search moves beyond lexical retrieval and improves grounding in retrieval-augmented generation. He discusses embedding quality, reranking, and application-specific evaluation through Continue.dev's codebase-search example and domain-specialized models, then highlights multimodal retrieval as an emerging direction."
  },
  {
    "slug": "the-state-of-generative-media-today",
    "title": "The State of Generative Media Today",
    "url": "https://ai.engineer/talks/the-state-of-generative-media-today",
    "videoId": "P370D8Kmlkw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1034000,
    "speakers": [
      {
        "slug": "gorkem-yurtseven",
        "name": "Gorkem Yurtseven",
        "organization": "Fal"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "fal co-founder Gorkem Yurtseven surveys the evolution of generative media, tracing computer-generated art and GANs through DALL·E 2, Midjourney, Stable Diffusion, SDXL, and FLUX. He discusses fal's inference platform and applications in advertising, interactive campaigns, e-commerce, and generative video, including Sora, before highlighting newer image-editing capabilities in FLUX.1 Kontext and GPT-4o."
  },
  {
    "slug": "the-state-of-mcp-observability-observable-tools-alex-volkov-and-benjamin-eckel-weights-biases-an",
    "title": "The State of MCP Observability: Observable.tools — Alex Volkov and Benjamin Eckel, Weights & Biases and Dylibso",
    "url": "https://ai.engineer/talks/the-state-of-mcp-observability-observable-tools-alex-volkov-and-benjamin-eckel-weights-biases-an",
    "videoId": "Lcqat4iP_lE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1016000,
    "speakers": [
      {
        "slug": "alex-volkov",
        "name": "Alex Volkov",
        "organization": "Weights & Biases"
      },
      {
        "slug": "benjamin-eckel",
        "name": "Benjamin Eckel",
        "organization": "Dylibso"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Alex Volkov of Weights & Biases and Benjamin Eckel of Dylibso explain how external Model Context Protocol tools create observability blind spots for production AI agents. They demonstrate W&B Weave's Python-based MCP tracing, introduce the Observable.tools initiative, and advocate vendor-neutral OpenTelemetry context propagation and distributed traces, including a working [REDACTED:url]run-to-Weave example."
  },
  {
    "slug": "the-state-of-model-routing-nvidia-cognition-openrouter",
    "title": "The State of Model Routing — NVIDIA, Cognition, OpenRouter",
    "url": "https://ai.engineer/talks/the-state-of-model-routing-nvidia-cognition-openrouter",
    "videoId": "QHBjufYK8TA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2897000,
    "speakers": [
      {
        "slug": "walden",
        "name": "Walden",
        "organization": "Cognition"
      },
      {
        "slug": "carter",
        "name": "Carter",
        "organization": "NVIDIA"
      },
      {
        "slug": "tanay",
        "name": "Tanay"
      },
      {
        "slug": "alex-atallah",
        "name": "Alex Atallah",
        "organization": "OpenRouter"
      },
      {
        "slug": "nav",
        "name": "Nav"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "A conference panel featuring participants from NVIDIA, Cognition, and OpenRouter examines how multi-model routing balances model capability, application quality, inference cost, and local hardware utilization. Discussion covers Cognition’s Devin Fusion, frontier versus smaller models, context compaction, evaluation and agent behavior, KV-cache-aware routing and GPU cache offloading, NVIDIA Dynamo and DGX Spark, prompt tuning, and collaborative agents."
  },
  {
    "slug": "the-unbearable-lightness-of-agent-optimization",
    "title": "The Unbearable Lightness of Agent Optimization",
    "url": "https://ai.engineer/talks/the-unbearable-lightness-of-agent-optimization",
    "videoId": "zfvEMNmVlNY",
    "event": "AI Engineer Code 2025",
    "durationMs": 1078000,
    "speakers": [
      {
        "slug": "alberto-romero",
        "name": "Alberto Romero",
        "organization": "Jointly"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Alberto Romero of Jointly presents Meta-ACE, a framework for optimizing AI agents across context, compute, verification, memory, and parameter adaptation. After reviewing Agentic Context Engineering and its vulnerabilities to flawed reflection, weak feedback, and context-only optimization, he describes task profiling, a learned meta-controller, quality gates, hierarchical verification, and confidence-weighted consensus across models including GPT-4, Claude, and DeepSeek. The presentation emphasizes resource-aware strategy allocation and applications to regulated industries and multimodal systems."
  },
  {
    "slug": "the-unofficial-guide-to-apple-s-private-cloud-compute",
    "title": "The Unofficial Guide to Apple’s Private Cloud Compute",
    "url": "https://ai.engineer/talks/the-unofficial-guide-to-apple-s-private-cloud-compute",
    "videoId": "CCsWZ5bJlO8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1236000,
    "speakers": [
      {
        "slug": "jonathan-mortensen",
        "name": "Jonathan Mortensen",
        "organization": "Confident Security"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Jonathan Mortensen of Confident Security presents an independent guide to Apple’s Private Cloud Compute, examining how remote AI inference can preserve user privacy without prohibitive dedicated-compute costs. He explains enforceable architectural requirements including stateless processing and non-targetability, describes how remote attestation and a public software transparency log establish verifiable trust, and discusses data-center identity checks alongside latency and compute-cost trade-offs. He explicitly states that he does not work for or represent Apple."
  },
  {
    "slug": "the-unreasonable-effectiveness-of-prompt-learning",
    "title": "The Unreasonable Effectiveness of Prompt Learning",
    "url": "https://ai.engineer/talks/the-unreasonable-effectiveness-of-prompt-learning",
    "videoId": "pP_dSNz_EdQ",
    "event": "AI Engineer Code 2025",
    "durationMs": 656000,
    "speakers": [
      {
        "slug": "aparna-dhinakaran",
        "name": "Aparna Dhinakaran",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Arize AI co-founder Aparna Dhinakaran describes an RL-inspired prompt-learning loop that improves coding agents by updating persistent instructions rather than fine-tuning model weights. Using Claude Code and Cline on a 150-example SWE-bench evaluation, she reports approximate initial issue-resolution rates of 40% and 30%, respectively. The workflow runs coding tasks and unit tests, asks an LLM judge to explain failures, and feeds those explanations into a meta-prompt that generates improved CLAUDE.md or Cline rules."
  },
  {
    "slug": "the-unreasonable-effectiveness-of-separating-the-task-from-the-model",
    "title": "The Unreasonable Effectiveness of Separating the Task from the Model",
    "url": "https://ai.engineer/talks/the-unreasonable-effectiveness-of-separating-the-task-from-the-model",
    "videoId": "GgLQ02aO-hs",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1031000,
    "speakers": [
      {
        "slug": "maxime-rivest",
        "name": "Maxime Rivest",
        "organization": "DSPy"
      },
      {
        "slug": "isaac-miller",
        "name": "Isaac Miller",
        "organization": "cmpnd"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Maxime Rivest and Isaac Miller explain how DSPy separates an AI task's input-output signature from its underlying model, prompt, reasoning strategy, and other implementation choices. They demonstrate composable Python programs, evaluation-guided optimization, conditional reasoning, enterprise cost reduction, and integration of recursive language models for long-context workloads. The presentation also explores using qualitative user feedback to generate evaluations and improve AI systems while preserving stable task contracts."
  },
  {
    "slug": "the-ux-of-ai-making-ai-powered-apps-your-users-don-t-hate",
    "title": "The UX of AI: Making AI-Powered Apps Your Users Don't Hate",
    "url": "https://ai.engineer/talks/the-ux-of-ai-making-ai-powered-apps-your-users-don-t-hate",
    "videoId": "L3RuP_q8Bwc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2159000,
    "speakers": [
      {
        "slug": "kathryn-grayson-nanz",
        "name": "Kathryn Grayson Nanz",
        "organization": "Progress Software"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Kathryn Grayson Nanz, Senior Design and Developer Advocate at Progress Software, explains how to design AI-powered applications around real user problems rather than assuming users understand prompts, hallucinations, or retrieval augmented generation. Drawing on the evolution of Macintosh interfaces, she recommends introducing unfamiliar capabilities gradually, making AI-generated changes and autonomous actions visible, clarifying verification requirements, access and costs, and providing practical next-step actions that help users apply generated outputs."
  },
  {
    "slug": "the-voice-first-ai-overlay-designing-conversational-co-pilots",
    "title": "The Voice-First AI Overlay: Designing Conversational Co-Pilots",
    "url": "https://ai.engineer/talks/the-voice-first-ai-overlay-designing-conversational-co-pilots",
    "videoId": "y9YQc9a3gNw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1004000,
    "speakers": [
      {
        "slug": "gregory-bruss",
        "name": "Gregory Bruss"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Gregory Bruss presents a voice-first AI overlay that assists human-to-human conversations without joining them as a third speaker. A language-learning demonstration uses caption scraping, debouncing, contextual LLM suggestions, and translation endpoints. The talk situates this approach alongside ambient agents and improving voice infrastructure, then examines cognitive load, interaction timing, round-trip latency, and cascading speech-recognition errors."
  },
  {
    "slug": "the-web-browser-is-all-you-need",
    "title": "The Web Browser Is All You Need",
    "url": "https://ai.engineer/talks/the-web-browser-is-all-you-need",
    "videoId": "YRGjll7uu5w",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1051000,
    "speakers": [
      {
        "slug": "paul-klein-iv",
        "name": "Paul Klein IV",
        "organization": "Browserbase"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Browserbase founder Paul Klein IV argues that headless web browsers provide a practical integration layer between AI agents and websites lacking dedicated APIs or MCP servers. He outlines browser-based MCP infrastructure, contrasts DOM-based, vision-based, and computer-use agents, demonstrates browser-agent development in Cursor, and addresses enterprise use cases, CAPTCHAs, compliance, robots.txt, and responsible automation during audience questions."
  },
  {
    "slug": "the-weekend-ai-engineer",
    "title": "The Weekend AI Engineer",
    "url": "https://ai.engineer/talks/the-weekend-ai-engineer",
    "videoId": "MarPORTD8vo",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1309000,
    "speakers": [
      {
        "slug": "hassan-el-mghari",
        "name": "Hassan El Mghari",
        "organization": "Vercel"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Hassan El Mghari explains how he built open-source AI applications on weekends and attracted millions of users. Examples include automated image descriptions, generated QR codes, article summaries, Git commit messages, photo restoration, and room redesign. He emphasizes readily available AI APIs, rapid experimentation, organic distribution through search and word of mouth, and consistent practice."
  },
  {
    "slug": "the-z-l-continuum-should-ai-engineers-still-read-code",
    "title": "The Z/L Continuum: Should AI Engineers Still Read Code?",
    "url": "https://ai.engineer/talks/the-z-l-continuum-should-ai-engineers-still-read-code",
    "videoId": "ZpK5PWX2YRM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1295000,
    "speakers": [
      {
        "slug": "alex-volkov",
        "name": "Alex Volkov",
        "organization": "W&B from CoreWeave"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Alex Volkov, host of ThursdAI, presents the Z/L Continuum, a task-specific framework for deciding how closely engineers should inspect AI-generated code. Contrasting Mario Zechner's emphasis on careful review with Ryan Lopopolo's agent-forward approach, he examines Claude Code, METR capability measurements, rising development throughput, and reliability risks. He argues that traces, evals, shadow mode, observability, rollback, and increasingly autonomous loops should shift oversight toward system-level verification without eliminating human judgment."
  },
  {
    "slug": "think-you-can-build-a-game-with-ai-think-again-the-new-games-are-just-being-invented",
    "title": "Think You Can Build a Game with AI? Think Again! The New Games Are Just Being Invented!",
    "url": "https://ai.engineer/talks/think-you-can-build-a-game-with-ai-think-again-the-new-games-are-just-being-invented",
    "videoId": "grdoOC1BT1s",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1080000,
    "speakers": [
      {
        "slug": "danielle-an",
        "name": "Danielle An",
        "organization": "Meta"
      },
      {
        "slug": "david-hoe",
        "name": "David Hoe",
        "organization": "Meta"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Meta’s Danielle An and David Hoe present an audience-controlled game instead of conventional slides to demonstrate AI-assisted game creation. They explain how generative tools lower barriers for creators while arguing that artistic cohesion, gameplay, and human taste still distinguish compelling games. Examples include using key art to anchor an LLM, rapidly prototyping and playtesting ideas, and building multiplayer experiences with runtime-LLM-driven NPCs."
  },
  {
    "slug": "thinking-deeper-in-gemini",
    "title": "Thinking Deeper in Gemini",
    "url": "https://ai.engineer/talks/thinking-deeper-in-gemini",
    "videoId": "8EQo4J2BWKw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1093000,
    "speakers": [
      {
        "slug": "jack-rae",
        "name": "Jack Rae",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Google DeepMind researcher Jack Rae explains Gemini Thinking as an approach to overcoming reasoning and intelligence bottlenecks through additional inference-time computation. He traces language-model progress from recurrent neural networks to contemporary conversational systems, discusses token processing, reasoning traces, and reinforcement learning, and highlights adaptive efficiency, multimodal mathematical reasoning, Gemini-assisted reconstruction of a DQN training setup, and future systems operating over millions of inference tokens."
  },
  {
    "slug": "thinner-agents-on-a-smarter-substrate-the-ontology-based-semantic-layer",
    "title": "Thinner Agents on a Smarter Substrate: The Ontology-based Semantic Layer",
    "url": "https://ai.engineer/talks/thinner-agents-on-a-smarter-substrate-the-ontology-based-semantic-layer",
    "videoId": "VGN22pPpb-8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 666000,
    "speakers": [
      {
        "slug": "emil-eifrem",
        "name": "Emil Eifrem",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "finance",
        "name": "Finance"
      }
    ],
    "summary": "Neo4j CEO Emil Eifrem argues that enterprise AI teams repeatedly rebuild brittle integrations across databases, warehouses, and identity services, creating discovery, trust, governance, and maintenance problems. Using a bank-account-opening agent as an example, he proposes thin agents operating on a shared graph-based semantic substrate that connects a business ontology to a technical ontology and source-system mappings. Agent execution feedback then informs future source selection and enables learning across agents."
  },
  {
    "slug": "this-video-was-edited-with-ai-agent-but-how",
    "title": "This video was edited with AI agent. But how?",
    "url": "https://ai.engineer/talks/this-video-was-edited-with-ai-agent-but-how",
    "videoId": "r0AG44qYKsI",
    "event": "AI Engineer Summit 2025",
    "durationMs": 300000,
    "speakers": [
      {
        "slug": "muhtesem",
        "name": "Muhtesem",
        "organization": "Re-skill"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Muhtesem presents a collaboration between Re-skill and Diffusion Studio on an open-source Python video-editing agent. The agent uses Playwright to operate an agent-oriented browser editor, generates JavaScript or TypeScript compositions with Diffusion Studio Core, renders through WebCodecs, and transfers files using the Chrome DevTools Protocol. Its workflow combines code-generating edits, retrieval-augmented documentation search, and sampled-frame visual feedback before rendering; llms.txt and GPU-accelerated remote browser sessions support discoverability and scaling."
  },
  {
    "slug": "through-the-ai-fog-the-architectural-decision-the-next-24-months-of-agentic-security-depends-on",
    "title": "Through the AI Fog: The architectural decision the next 24 months of agentic security depends on.",
    "url": "https://ai.engineer/talks/through-the-ai-fog-the-architectural-decision-the-next-24-months-of-agentic-security-depends-on",
    "videoId": "1EZdpEhwmNc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1409000,
    "speakers": [
      {
        "slug": "manoj-nair",
        "name": "Manoj Nair",
        "organization": "Snyk"
      },
      {
        "slug": "ezra",
        "name": "Ezra"
      },
      {
        "slug": "randall",
        "name": "Randall"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Snyk executive Manoj Nair discusses security architecture for autonomous AI development, emphasizing risks from poisoned agent skills, MCP-connected tools, chained exploits, model-dependent sensitive-data leakage, and growing remediation demands. A guest demonstrator, identified only as Ezra, presents security tooling before connectivity problems force a switch to a recorded demonstration. Nair closes by introducing Evo by Snyk and advocating stronger tools and training for AI security engineers."
  },
  {
    "slug": "tldraw-computer",
    "title": "tldraw computer",
    "url": "https://ai.engineer/talks/tldraw-computer",
    "videoId": "1C2TdPkj6aQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1125000,
    "speakers": [
      {
        "slug": "steve-ruiz",
        "name": "Steve Ruiz",
        "organization": "tldraw"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "tldraw founder Steve Ruiz demonstrates how its React-based infinite-canvas SDK supports embedded web applications and AI-powered visual programming. He presents vision-model and stop-motion experiments, then shows tldraw computer executing graphs that transform connected inputs into text, images, and speech, including cyclic workflows and potential customer-email sentiment automation."
  },
  {
    "slug": "tlms-tiny-llms-and-agents-on-edge-devices-with-litert-lm",
    "title": "TLMs: Tiny LLMs and Agents on Edge Devices with LiteRT-LM",
    "url": "https://ai.engineer/talks/tlms-tiny-llms-and-agents-on-edge-devices-with-litert-lm",
    "videoId": "BKWpYIWvAo4",
    "event": "AI Engineer Europe 2026",
    "durationMs": 4858000,
    "speakers": [
      {
        "slug": "cormac-brick",
        "name": "Cormac Brick",
        "organization": "Google"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Google AI Edge technical lead Cormac Brick demonstrates how LiteRT-LM and Gemma models enable agents and language-model applications on phones and other constrained devices. The workshop covers Android and iOS deployment, acceleration across Intel, Raspberry Pi, and Qualcomm platforms, practical agent-skill development with Gemini CLI, small-model guardrails, Apple's FastVLM, tiny-model fine-tuning, modular model pipelines, and memory constraints."
  },
  {
    "slug": "to-the-moon-navigating-deep-context-in-legacy-code-with-augment-agent",
    "title": "To the moon! Navigating deep context in legacy code with Augment Agent",
    "url": "https://ai.engineer/talks/to-the-moon-navigating-deep-context-in-legacy-code-with-augment-agent",
    "videoId": "6NIr_cYPglk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 924000,
    "speakers": [
      {
        "slug": "forrest-brazeal",
        "name": "Forrest Brazeal",
        "organization": "Freeman & Forrest"
      },
      {
        "slug": "matt-ball",
        "name": "Matt Ball",
        "organization": "Augment Code"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Forrest Brazeal and Augment Code solutions architect Matt Ball demonstrate how Augment Agent investigates and modernizes unfamiliar legacy software using the Apollo 11 Guidance Computer assembly code. They discuss existing-IDE integration and enterprise security, trace the 1202 scheduler alarm, distinguish the P65 automatic and P66 manually controlled lunar-landing routines, and use the agent's autonomous mode to write missing simulator code and complete a simulated landing. The presentation is a condensed Expo Sessions version of an earlier, separate workshop."
  },
  {
    "slug": "tool-calling-is-not-just-plumbing-for-ai-agents",
    "title": "Tool Calling Is Not Just Plumbing for AI Agents",
    "url": "https://ai.engineer/talks/tool-calling-is-not-just-plumbing-for-ai-agents",
    "videoId": "zuMw0pkPXpU",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1518000,
    "speakers": [
      {
        "slug": "roy-derks",
        "name": "Roy Derks",
        "organization": "IBM"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "IBM developer Roy Derks argues that AI-agent tools deserve the same deliberate engineering as agents themselves. He compares traditional and meta tool calling, explains effective tool descriptions, typed arguments and structured outputs, and shows how external APIs, databases, search and computation can extend language models. He advocates framework-independent tool platforms interoperable with LangChain, LangGraph, CrewAI and AutoGen, and describes dynamically generating tools from GraphQL or SQL schemas."
  },
  {
    "slug": "tools-for-the-next-generation-of-ai-engineers",
    "title": "Tools for the Next Generation of AI Engineers",
    "url": "https://ai.engineer/talks/tools-for-the-next-generation-of-ai-engineers",
    "videoId": "ySYLsoAhXmg",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1307000,
    "speakers": [
      {
        "slug": "stefania-druga",
        "name": "Stefania Druga",
        "organization": "Google"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Stefania Druga argues that children should help design AI tools rather than merely consume them. She demonstrates how Scratch and Cognimates introduce visual programming, robotics, and image/text model training, then explores AI-assisted creative coding, student control over suggestions, collaborative game-building agents, and Scratch-based optical character recognition."
  },
  {
    "slug": "trading-desks-to-clinical-trials-parallels-in-applied-vertical-ai",
    "title": "Trading Desks to Clinical Trials: Parallels in Applied Vertical AI",
    "url": "https://ai.engineer/talks/trading-desks-to-clinical-trials-parallels-in-applied-vertical-ai",
    "videoId": "Yphdry8ttAQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1202000,
    "speakers": [
      {
        "slug": "ayush-bhardwaj",
        "name": "Ayush Bhardwaj",
        "organization": "Allos AI"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Allos AI Tech Lead Ayush Bhardwaj compares building specialized AI systems for hedge funds and pharmaceutical research. He argues that effective vertical AI requires narrowly scoped agents, proprietary domain data, observability, measurable business returns, and expert judgment capable of evaluating specialized outputs. Because models and agent infrastructure are widely available, he identifies domain expertise, hard-to-obtain data, and practical error analysis as more durable advantages."
  },
  {
    "slug": "training-agentic-reasoners",
    "title": "Training Agentic Reasoners",
    "url": "https://ai.engineer/talks/training-agentic-reasoners",
    "videoId": "PbHm2qKnu10",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1157000,
    "speakers": [
      {
        "slug": "will-brown",
        "name": "Will Brown",
        "organization": "Prime Intellect"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Prime Intellect research lead Will Brown argues that language-model reasoning and tool-using agents share the same reinforcement-learning foundations. He discusses scaling RL, the complexity of the verl framework, GRPO and DeepSeekMath, agent harnesses and environments, evaluation of ambiguous tasks, multi-turn interaction, and supervised-fine-tuning warm-ups that make smaller-model training more accessible."
  },
  {
    "slug": "training-albatross-an-expert-finance-llm",
    "title": "Training Albatross: An Expert Finance LLM",
    "url": "https://ai.engineer/talks/training-albatross-an-expert-finance-llm",
    "videoId": "of-SV35YqvY",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 980000,
    "speakers": [
      {
        "slug": "leo-pekelis",
        "name": "Leo Pekelis",
        "organization": "Gradient"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Gradient chief scientist Leo Pekelis explains how the Albatross finance-domain language model adapts a Llama 2 foundation using domain-specific training data, human review, and synthetic augmentation. He contrasts long-context, in-context learning with brittle retrieval pipelines for reducing hallucinations and reports improvements on finance-specific evaluations and NVIDIA's RULER long-context benchmark."
  },
  {
    "slug": "training-an-llm-from-scratch-locally",
    "title": "Training an LLM from Scratch, Locally",
    "url": "https://ai.engineer/talks/training-an-llm-from-scratch-locally",
    "videoId": "UsB70Tf5zcE",
    "event": "AI Engineer Europe 2026",
    "durationMs": 4886000,
    "speakers": [
      {
        "slug": "angelos-perivolaropoulos",
        "name": "Angelos Perivolaropoulos",
        "organization": "ElevenLabs"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "ElevenLabs research engineer Angelos Perivolaropoulos leads a hands-on workshop on training a small GPT-2-style language model locally from scratch using Torch. He introduces ElevenLabs’ Scribe v2, then covers tokenizer design, transformer architecture, nanoGPT-inspired implementation, optimization schedules, validation and overfitting, initial loss behavior, generated outputs, and an interactive participant challenge with audience questions."
  },
  {
    "slug": "training-frontier-models-to-out-think-hackers-uri-rolls-arithmetic-thom-wolf-hugging-face",
    "title": "Training Frontier Models to Out-Think Hackers — Uri Rolls, Arithmetic & Thom Wolf, Hugging Face",
    "url": "https://ai.engineer/talks/training-frontier-models-to-out-think-hackers-uri-rolls-arithmetic-thom-wolf-hugging-face",
    "videoId": "O-CBZ3JtRvo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1048000,
    "speakers": [
      {
        "slug": "uri-rolls",
        "name": "Uri Rolls",
        "organization": "Arithmetic"
      },
      {
        "slug": "thom-wolf",
        "name": "Thom Wolf",
        "organization": "Hugging Face"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Uri Rolls of Arithmetic and Hugging Face cofounder Thom Wolf present an interactive cybersecurity benchmark focused on broken access control, arguing that effective AI defenders must reason through changing systems rather than merely identify isolated vulnerabilities. They describe researcher-discovered zero days, deterministic grading, an authorization flaw caused by inconsistent name-versus-ID checks, and difficult evaluations involving GPT-5.5 and Opus, while advocating capable open-source models for scalable defense."
  },
  {
    "slug": "training-krea-2-what-matters-in-generative-model-training",
    "title": "Training Krea 2 - What matters in generative model training.",
    "url": "https://ai.engineer/talks/training-krea-2-what-matters-in-generative-model-training",
    "videoId": "-tviRdpmHvs",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1306000,
    "speakers": [
      {
        "slug": "sangwu-lee",
        "name": "Sangwu Lee",
        "organization": "Krea.ai"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Krea researcher Sangwu Lee explains how Krea 2 was trained as an image foundation model emphasizing fast creative exploration, aesthetic diversity, and broad visual world knowledge. He describes avoiding synthetic training images, preserving unconventional styles, combining OCR with vision-language-model captioning, deduplicating and filtering datasets with pHash, MD5, classifiers, and sparse autoencoders, and using Wikipedia concept rankings to improve coverage."
  },
  {
    "slug": "trends-across-the-ai-frontier",
    "title": "Trends Across the AI Frontier",
    "url": "https://ai.engineer/talks/trends-across-the-ai-frontier",
    "videoId": "sRpqPgKeXNk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1072000,
    "speakers": [
      {
        "slug": "micah-hill-smith",
        "name": "Micah Hill-Smith",
        "organization": "Artificial Analysis"
      },
      {
        "slug": "george-cameron",
        "name": "George Cameron",
        "organization": "Artificial Analysis"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Artificial Analysis co-founder George Cameron argues that AI progress has several frontiers beyond benchmark intelligence, including reasoning efficiency, latency, openness, and compute demand. He compares leading models, reports that intelligence-index evaluations consumed approximately 7 million tokens for GPT-4.1, 72 million for high-reasoning o4-mini, and 130 million for Gemini 2.5 Pro, and contrasts GPT-4.1’s 4.7-second median response with more than 40 seconds for o4-mini-high. Cameron also describes the narrowing gap between open-weight and proprietary models and concludes that sequential model requests will continue increasing demand for compute."
  },
  {
    "slug": "trunk-tools-launch-disrupting-the-15-trillion-construction-industry-with-autonomous-agents",
    "title": "Trunk Tools Launch: Disrupting the $15 Trillion Construction Industry with Autonomous Agents",
    "url": "https://ai.engineer/talks/trunk-tools-launch-disrupting-the-15-trillion-construction-industry-with-autonomous-agents",
    "videoId": "L0kBWyziFlc",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 332000,
    "speakers": [
      {
        "slug": "dr-sarah-buchner",
        "name": "Dr. Sarah Buchner",
        "organization": "Trunk Tools"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Dr. Sarah Buchner explains how Trunk Tools applies construction-specific retrieval-augmented generation to massive collections of project documents, citing a skyscraper project with 3.6 million pages. She demonstrates source-grounded question answering and argues that the larger opportunity lies in human-centered vertical agents that detect discrepancies, generate RFIs, and reduce expensive construction rework."
  },
  {
    "slug": "trust-but-verify",
    "title": "Trust, but Verify",
    "url": "https://ai.engineer/talks/trust-but-verify",
    "videoId": "9-vGxMoUM9Y",
    "event": "AI Engineer Summit 2023",
    "durationMs": 1181000,
    "speakers": [
      {
        "slug": "shreya-rajpal",
        "name": "Shreya Rajpal",
        "organization": "Guardrails AI"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Guardrails AI co-founder and CEO Shreya Rajpal explains why generative AI prototypes often fail in production: unlike deterministic software APIs, LLMs produce stochastic outputs and can violate structural, correctness, or security requirements. She proposes a trust-but-verify programming model in which Guardrails wraps LLM calls with configurable validation, constraint checking, and rule-based heuristics before outputs are accepted."
  },
  {
    "slug": "trust-but-verify-high-fidelity-reasoning-in-agentic-workflows",
    "title": "Trust, but Verify: High-Fidelity Reasoning in Agentic Workflows",
    "url": "https://ai.engineer/talks/trust-but-verify-high-fidelity-reasoning-in-agentic-workflows",
    "videoId": "MWTJIAwAAnk",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1270000,
    "speakers": [
      {
        "slug": "mike-conover",
        "name": "Mike Conover",
        "organization": "Brightwave"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Brightwave founder and CEO Mike Conover explains how knowledge agents can accelerate investment research and due diligence across large collections of filings, transcripts, contracts, and other financial documents. Drawing on his Databricks Dolly background, he examines compounding extraction errors, reinforcement learning over tool-use sequences, decomposed research tasks, factual-entailment checks, and model self-correction. He argues that trustworthy financial research products need transparent reasoning and interactive, interrogable findings rather than relying on chat interfaces or unverified reports alone."
  },
  {
    "slug": "turn-10-994-notes-into-your-agents-memory",
    "title": "Turn 10,994 Notes Into Your Agents' Memory",
    "url": "https://ai.engineer/talks/turn-10-994-notes-into-your-agents-memory",
    "videoId": "ZRM_TfEZcIo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 2372000,
    "speakers": [
      {
        "slug": "paul-iusztin",
        "name": "Paul Iusztin",
        "organization": "Decoding AI"
      },
      {
        "slug": "louis-francois-bouchard",
        "name": "Louis-François Bouchard",
        "organization": "Towards AI"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Paul Iusztin and Louis-François Bouchard demonstrate AI Research OS, an open-source, file-based research-memory layer that converts notes and other personal knowledge sources into persistent context for AI agents. They compare one-off tools such as ChatGPT, Codex, Claude Code, and NotebookLM with reusable deep-research workflows, explain lightweight personal research wikis, demonstrate the workshop repository, and close with an Agent Engineering course."
  },
  {
    "slug": "turning-fails-into-features-zapier-s-hard-won-eval-lessons",
    "title": "Turning Fails into Features: Zapier’s Hard-Won Eval Lessons",
    "url": "https://ai.engineer/talks/turning-fails-into-features-zapier-s-hard-won-eval-lessons",
    "videoId": "blrovBxxN9o",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 975000,
    "speakers": [
      {
        "slug": "rafal-wilinski",
        "name": "Rafal Wilinski",
        "organization": "Zapier"
      },
      {
        "slug": "vitor-balocco",
        "name": "Vitor Balocco",
        "organization": "Zapier"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Rafal Wilinski and Vitor Balocco explain how Zapier improves AI agents by instrumenting complete execution traces, collecting explicit and implicit user-feedback signals, clustering frustrations, and connecting evaluation to meaningful business metrics. They demonstrate using Braintrust MCP and a reasoning model to investigate regressions and compare Gemini Pro with Claude, emphasizing that evaluations should account for intermediate tool calls, generated artifacts, differing model behavior, and subtle grading bias."
  },
  {
    "slug": "two-roads-to-durable-agents-replay-vs-snapshot-eric-allam-co-founder-trigger-dev",
    "title": "Two Roads to Durable Agents: Replay vs. Snapshot — Eric Allam, Co-founder, Trigger.dev",
    "url": "https://ai.engineer/talks/two-roads-to-durable-agents-replay-vs-snapshot-eric-allam-co-founder-trigger-dev",
    "videoId": "svCnShDvgQg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 996000,
    "speakers": [
      {
        "slug": "eric-allam",
        "name": "Eric Allam",
        "organization": "Trigger.dev"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Trigger.dev co-founder Eric Allam contrasts replay-based durable execution with snapshot-based infrastructure for long-running AI agents. He traces backend architecture from CGI and PHP through asynchronous workflow engines, then argues that agent durability requires separately preserving LLM conversation context and stateful execution environments containing files, memory, subprocesses, and tools. After comparing historical checkpointing and CRIU with Firecracker microVM snapshots, he reports sub-second snapshot creation, restores within a few hundred milliseconds, and approximately 15,000 virtual-machine starts per minute."
  },
  {
    "slug": "under-5-minutes-to-a-deployed-llm-endpoint-audry-hsu-runpod",
    "title": "Under 5 minutes to a deployed LLM endpoint — Audry Hsu, RunPod",
    "url": "https://ai.engineer/talks/under-5-minutes-to-a-deployed-llm-endpoint-audry-hsu-runpod",
    "videoId": "ILdE7FaAjVA",
    "event": "AI Engineer Europe 2026",
    "durationMs": 806000,
    "speakers": [
      {
        "slug": "audry-hsu",
        "name": "Audry Hsu",
        "organization": "RunPod"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Audry Hsu introduces RunPod’s GPU cloud infrastructure and describes its community-driven origins before comparing Pods, autoscaling Serverless workers, multi-node Clusters, and the repository-backed Hub. A console demonstration shows how to select a Hub listing, configure H100 GPUs with A100 fallback and worker scaling, and deploy an LLM endpoint. Hsu also previews a separate session on the Flash Python SDK for running functions remotely on GPUs."
  },
  {
    "slug": "understanding-ai-stakes-to-break-production-code",
    "title": "Understanding AI Stakes to Break Production Code",
    "url": "https://ai.engineer/talks/understanding-ai-stakes-to-break-production-code",
    "videoId": "gWLVu9OGvBg",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1404000,
    "speakers": [
      {
        "slug": "philip-rathle",
        "name": "Philip Rathle",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Neo4j CTO Philip Rathle facilitates an interactive roundtable on why promising generative AI prototypes struggle to reach production and how deployment requirements change with the consequences of incorrect answers. Participants discuss vector-based RAG versus knowledge graphs and GraphRAG, customer resistance to semantic-search interfaces, prioritizing production-ready projects, model cost and performance tradeoffs, transparent AI interaction design, and human oversight in regulated behavioral-health applications."
  },
  {
    "slug": "understanding-is-the-new-bottleneck",
    "title": "Understanding is the new bottleneck",
    "url": "https://ai.engineer/talks/understanding-is-the-new-bottleneck",
    "videoId": "WkBPX-oDMnA",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1173000,
    "speakers": [
      {
        "slug": "geoffrey-litt",
        "name": "Geoffrey Litt",
        "organization": "Notion"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Notion design engineer Geoffrey Litt argues that human understanding, rather than code generation or correctness checking, is the emerging bottleneck in agent-assisted software development. He describes cognitive debt and demonstrates educationally informed practices including collaborative code explainers, interactive Notion HTML artifacts, comprehension checks, the Explain Diff skill, and Claude-generated microworlds for understanding software behavior."
  },
  {
    "slug": "unlock-agent-autonomy-the-runtime-for-ai-native-systems",
    "title": "Unlock Agent Autonomy: The Runtime for AI-Native Systems",
    "url": "https://ai.engineer/talks/unlock-agent-autonomy-the-runtime-for-ai-native-systems",
    "videoId": "zaGyGgLW3SM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1370000,
    "speakers": [
      {
        "slug": "tushar-jain",
        "name": "Tushar Jain",
        "organization": "Docker"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Docker engineering executive Tushar Jain argues that agent autonomy is limited less by model intelligence than by the ability to enforce safety at runtime. Using an agent that unexpectedly published a private report and an incident investigation that progressively requests GitHub and Slack access, he outlines containment outside the agent trust boundary, task-scoped capabilities, just-in-time MCP tools, and intent-aware policy enforcement. Demonstrations extend these controls to cloud execution, orchestrated agents, and Docker Sandboxes through the sbx CLI."
  },
  {
    "slug": "unlocking-africa-s-potential-with-ai-thabang-ledwaba",
    "title": "Unlocking Africa's Potential with AI — Thabang Ledwaba",
    "url": "https://ai.engineer/talks/unlocking-africa-s-potential-with-ai-thabang-ledwaba",
    "videoId": "eD_6jP1fkKs",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1577000,
    "speakers": [
      {
        "slug": "thabang-ledwaba",
        "name": "Thabang Ledwaba",
        "organization": "BBD Software"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Thabang Ledwaba of BBD Software argues that Africa should shape its own AI future through locally grounded, practical innovation. He contrasts generative and agentic AI, cites M-PESA as evidence of [REDACTED:origin] fintech ingenuity, critiques overengineered responses to [REDACTED:origin] public-service queues, and emphasizes simplicity, resourcefulness, ubuntu, and meaningful participation in the global technology economy."
  },
  {
    "slug": "unlocking-ai-powered-devops-within-your-organization",
    "title": "Unlocking AI Powered DevOps Within Your Organization",
    "url": "https://ai.engineer/talks/unlocking-ai-powered-devops-within-your-organization",
    "videoId": "C1NivhYS1sI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1333000,
    "speakers": [
      {
        "slug": "jon-peck",
        "name": "Jon Peck",
        "organization": "GitHub"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "GitHub developer advocate Jon Peck explains how organizations can adopt AI across DevOps by prioritizing useful efficiency gains, integrating assistants directly into development environments, choosing models appropriate to each task, and applying GitHub Copilot to existing codebases. He discusses agent-assisted project scoping, team learning and governance, increasing autonomy, and repository-level Model Context Protocol connections for external tools."
  },
  {
    "slug": "unlocking-developer-productivity-across-cpu-and-gpu-with-max",
    "title": "Unlocking Developer Productivity across CPU and GPU with MAX",
    "url": "https://ai.engineer/talks/unlocking-developer-productivity-across-cpu-and-gpu-with-max",
    "videoId": "7TnkqfX84gI",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1113000,
    "speakers": [
      {
        "slug": "chris-lattner",
        "name": "Chris Lattner",
        "organization": "Modular"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Modular co-founder Chris Lattner argues that fragmented inference frameworks impede secure, customizable production AI. He introduces MAX and Mojo as a unified, Python-oriented stack intended to simplify model deployment and CPU/GPU programming, drawing on established compiler and machine-learning infrastructure while targeting higher developer productivity and hardware performance."
  },
  {
    "slug": "unveiling-the-latest-gemma-model-advancements",
    "title": "Unveiling the latest Gemma model advancements",
    "url": "https://ai.engineer/talks/unveiling-the-latest-gemma-model-advancements",
    "videoId": "Xmkl27AM2VQ",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 985000,
    "speakers": [
      {
        "slug": "kathleen-kenealy",
        "name": "Kathleen Kenealy",
        "organization": "DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Google DeepMind research engineer and Gemma technical lead Kathleen Kenealy introduces the Gemma open-model family and announces Gemma 2 on its June 27, 2024 launch day. She describes responsible development and safety evaluations, deployment across TPUs, GPUs, local devices, and multiple ML frameworks, and specialized variants including CodeGemma, RecurrentGemma, and PaliGemma. The talk also discusses practical model evaluation and invites developers to contribute examples to the Gemma Cookbook."
  },
  {
    "slug": "useful-general-intelligence",
    "title": "Useful General Intelligence",
    "url": "https://ai.engineer/talks/useful-general-intelligence",
    "videoId": "Dj0b_cEBHBI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1198000,
    "speakers": [
      {
        "slug": "danielle-perszyk",
        "name": "Danielle Perszyk",
        "organization": "Amazon AGI SF Lab"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Amazon AGI SF Lab cognitive scientist Danielle Perszyk argues that useful general intelligence should augment human judgment and agency rather than reproduce or replace human intelligence. Drawing on predictive cognition, controlled hallucinations, Douglas Engelbart, and technosocial co-evolution, she frames reliable, aligned computer-use agents as human-centered tools. A Nova Act demonstration combines browser automation, Pydantic-based JSON extraction, Google Maps, and pandas to evaluate apartment listings and commute times, before concluding that useful products and human-agent interaction data can reinforce each other."
  },
  {
    "slug": "user-signal-die-at-the-retrieval-boundary",
    "title": "User Signal Die at the Retrieval Boundary",
    "url": "https://ai.engineer/talks/user-signal-die-at-the-retrieval-boundary",
    "videoId": "Jx4ZFEAq6bY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 937000,
    "speakers": [
      {
        "slug": "sonam-pankaj",
        "name": "Sonam Pankaj",
        "organization": "StarlightSearch Inc"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Starlight Search CEO Sonam Pankaj explains why ReAct-style agents repeatedly fail when observability traces and evaluation outcomes never influence future retrieval. She presents Reflect, an outcome-aware runtime memory layer that ranks prior experiences by semantic relevance and observed utility, converts successful lessons into reusable skills, reports company benchmark improvements, and demonstrates failure feedback with a SQL product-search agent."
  },
  {
    "slug": "using-agents-to-build-an-agent-company",
    "title": "Using agents to build an agent company",
    "url": "https://ai.engineer/talks/using-agents-to-build-an-agent-company",
    "videoId": "Dc99-zTMyMg",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 826000,
    "speakers": [
      {
        "slug": "joao-moura",
        "name": "João Moura",
        "organization": "crewAI"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "crewAI founder and CEO João Moura explains how autonomous, role-specialized AI agents evolved from automating his LinkedIn writing into a framework and company built around multi-agent orchestration. He discusses production requirements including tools, caching, memory, and training; demonstrates marketing and lead-qualification crews; and describes CLI-based agent training and automated deployment through GitHub and CrewAI Plus."
  },
  {
    "slug": "using-ai-to-build-an-infinite-game",
    "title": "Using AI to Build an Infinite Game",
    "url": "https://ai.engineer/talks/using-ai-to-build-an-infinite-game",
    "videoId": "_KFbT6eph5A",
    "event": "AI Engineer Summit 2023",
    "durationMs": 668000,
    "speakers": [
      {
        "slug": "jeff-schomay",
        "name": "Jeff Schomay"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "Jeff Schomay demonstrates Lost, a narrative exploration game that generates new scenes and visuals dynamically. He fine-tunes an OpenAI model on approximately 50 examples to produce consistent JSON scene definitions from shorter prompts, develops a custom Leonardo.Ai image model for coherent visual style, and connects both through an asset-server pipeline that validates generated JSON before requesting images. A gameplay demonstration highlights the resulting variety, while closing remarks identify 512-pixel image resolution and potential AI upscaling as improvement areas."
  },
  {
    "slug": "using-llms-to-secure-source-code",
    "title": "Using LLMs to Secure Source Code",
    "url": "https://ai.engineer/talks/using-llms-to-secure-source-code",
    "videoId": "imFedndyXYQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1290000,
    "speakers": [
      {
        "slug": "eugene-yan",
        "name": "Eugene Yan",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Anthropic's Eugene Yan explains how security teams can use LLMs and agentic harnesses to find, verify, prioritize, and patch real source-code vulnerabilities. Using Mozilla's Firefox work with Claude Mythos Preview and a Python SQL-injection example, he outlines a six-step workflow built around threat modeling, isolated sandboxes, discovery, verification, triage, and patching. He emphasizes system context, reproducible exploit testing, scarce engineering attention, and feedback loops that improve patch quality."
  },
  {
    "slug": "using-oss-models-to-build-ai-apps-with-millions-of-users",
    "title": "Using OSS models to build AI apps with millions of users",
    "url": "https://ai.engineer/talks/using-oss-models-to-build-ai-apps-with-millions-of-users",
    "videoId": "gcseUQJ6Gbg",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1127000,
    "speakers": [
      {
        "slug": "hassan-el-mghari",
        "name": "Hassan El Mghari",
        "organization": "Together AI"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Together AI developer-relations lead Hassan El Mghari explains how he rapidly builds, launches, and markets full-stack AI applications using open-source models. He demonstrates prompt-to-app generation and other lightweight products, describes a stack combining model inference, React, TypeScript, Postgres, Clerk, Prisma, storage, and analytics, and discusses prototyping, authentication, cost limits, timely model adoption, and audience questions."
  },
  {
    "slug": "using-rl-based-agent-to-detect-and-remediate-etl-pipeline-failures",
    "title": "Using RL-based Agent to Detect and Remediate ETL Pipeline Failures",
    "url": "https://ai.engineer/talks/using-rl-based-agent-to-detect-and-remediate-etl-pipeline-failures",
    "videoId": "LrGCT7G_rU8",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 881000,
    "speakers": [
      {
        "slug": "anna-marie-benzon",
        "name": "Anna Marie Benzon"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Anna Marie Benzon presents an event-driven AWS architecture that detects ETL failures and combines deterministic anomaly rules, interpretable Q-learning, bounded remediation actions, and independent safety overrides. AWS Glue failure events flow through EventBridge and Lambda, using CloudWatch logs and Glue Data Catalog metadata to choose actions such as retry, schema coercion, rollback, quarantine, or human escalation. In a controlled synthetic benchmark, the anomaly detector achieved precision 1.0, recall 0.8, and F1 0.889; successfully resolved cases averaged approximately 5.24 minutes versus a modeled 2.5-day manual baseline. Ablation results found the RL policy matched an equivalent deterministic policy, emphasizing structured state and safety constraints rather than claiming an independent RL performance advantage. A sanitized public repository supplies reproducibility materials while excluding client-sensitive information."
  },
  {
    "slug": "using-spec-driven-development-for-production-workflows",
    "title": "Using Spec-Driven Development for Production Workflows",
    "url": "https://ai.engineer/talks/using-spec-driven-development-for-production-workflows",
    "videoId": "IddXPepIAS4",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1067000,
    "speakers": [
      {
        "slug": "redacted-username",
        "name": "[REDACTED:username]",
        "organization": "Amazon Web Services"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "AWS senior developer advocate [REDACTED:username] explains how spec-driven development improves AI-assisted production workflows by creating and reviewing requirements, design documents, and implementation plans before code is generated. He discusses instruction files, Kiro steering documents and skills, human accountability for code review, Kiro's IDE and CLI spec workflows, and MCP integrations with project-management tools alongside security considerations."
  },
  {
    "slug": "ux-design-principles-for-semi-autonomous-multi-agent-systems",
    "title": "UX Design Principles for (Semi) Autonomous Multi-Agent Systems",
    "url": "https://ai.engineer/talks/ux-design-principles-for-semi-autonomous-multi-agent-systems",
    "videoId": "fmZWvE7yDZo",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1228000,
    "speakers": [
      {
        "slug": "victor-dibia",
        "name": "Victor Dibia",
        "organization": "Microsoft Research"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Microsoft Research engineer Victor Dibia presents human-centered UX principles for semi-autonomous multi-agent systems, drawing on GitHub Copilot, LIDA, AutoGen, and AutoGen Studio. He demonstrates BlenderLM, a from-scratch agentic application that converts natural-language requests into Blender 3D operations, and emphasizes capability discovery, observable agent behavior, cost-aware delegation, and interruptibility through checkpointing, rollback, pause, and resume."
  },
  {
    "slug": "vector-search-benchmark-eting",
    "title": "Vector Search Benchmark[eting]",
    "url": "https://ai.engineer/talks/vector-search-benchmark-eting",
    "videoId": "YrUBFXa1KUY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 850000,
    "speakers": [
      {
        "slug": "philipp-krenn",
        "name": "Philipp Krenn",
        "organization": "Elastic"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Elastic developer-relations leader Philipp Krenn critiques misleading vector-search vendor benchmarks and recommends testing representative, application-specific workloads. He highlights read-only benchmark bias, counterintuitive HNSW filtering costs, and gradual performance regressions, then introduces Elasticsearch Rally and configurable tracks for reproducible, tailored benchmarking."
  },
  {
    "slug": "velocity-sickness-what-happens-when-your-whole-team-gets-10x-faster",
    "title": "Velocity Sickness: What Happens When Your Whole Team Gets 10x Faster",
    "url": "https://ai.engineer/talks/velocity-sickness-what-happens-when-your-whole-team-gets-10x-faster",
    "videoId": "Kz4QJmNrVXU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1237000,
    "speakers": [
      {
        "slug": "matt-dailey",
        "name": "Matt Dailey",
        "organization": "Ref."
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Ref. founder and CEO Matt Dailey describes velocity sickness: AI makes individual engineers dramatically faster while overwhelming teams with pull requests, fragmented initiatives, abandoned agent sessions, and opaque automated decisions. He argues that implementation-oriented IDEs and ephemeral chats no longer match engineering work, and recommends preserving human ownership through shared, durable decision documents and collaborative planning that redirect speed from code output toward coherent ideas and meaningful product impact."
  },
  {
    "slug": "vending-bench-long-horizon-agent-evals-lukas-petersson-andon-labs",
    "title": "Vending-Bench: Long-Horizon Agent Evals — Lukas Petersson, Andon Labs",
    "url": "https://ai.engineer/talks/vending-bench-long-horizon-agent-evals-lukas-petersson-andon-labs",
    "videoId": "cO8qC6HBuBg",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1085000,
    "speakers": [
      {
        "slug": "lukas-petersson",
        "name": "Lukas Petersson",
        "organization": "Andon Labs"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Andon Labs co-founder Lukas Petersson describes Vending-Bench, a long-horizon evaluation in which AI agents operate simulated vending businesses and compete in a multiplayer arena. He reports emergent collusion, deception, and power-seeking behavior, then explains why simulation awareness motivates real-world deployments including a San Francisco shop, a Stockholm café, and AI radio stations. The talk compares model performance and alignment, describes replacing Gemini with GPT at the café, and examines how replaying incidents across models can improve the reproducibility of real-world agent evaluations."
  },
  {
    "slug": "veo-3-for-developers",
    "title": "Veo 3 for developers",
    "url": "https://ai.engineer/talks/veo-3-for-developers",
    "videoId": "hlcAZ2lX_ZI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1237000,
    "speakers": [
      {
        "slug": "paige-bailey",
        "name": "Paige Bailey",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "Google DeepMind developer-relations engineering lead Paige Bailey demonstrates Veo 3 and related generative-media models, contrasting Veo 2 editing and API workflows with video generation that incorporates audio, dialogue, lip synchronization, and image-to-video animation. She discusses Imagen 4, Lyria 2, Music AI Sandbox, MusicFX, SynthID watermarking, earlier video models, and a practical Gemini-orchestrated workflow that segments prompts and adds generated music."
  },
  {
    "slug": "vercel-ai-sdk-masterclass-from-fundamentals-to-deep-research",
    "title": "Vercel AI SDK Masterclass: From Fundamentals to Deep Research",
    "url": "https://ai.engineer/talks/vercel-ai-sdk-masterclass-from-fundamentals-to-deep-research",
    "videoId": "kDlqpN1JyIw",
    "event": "AI Engineer Summit 2025",
    "durationMs": 3592000,
    "speakers": [
      {
        "slug": "nico-albanese",
        "name": "Nico Albanese",
        "organization": "Vercel"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Vercel’s Nico Albanese introduces AI SDK fundamentals, including local Node.js/TypeScript setup, generateText, GPT-4o mini, alternative providers, tool-assisted web access, and structured outputs. He then builds an agentic deep-research workflow that generates search queries, evaluates results, recursively pursues follow-up questions, maintains accumulated research state, and produces a Markdown report."
  },
  {
    "slug": "verifiable-environments-for-ai-in-biology-kenny-workman-latchbio",
    "title": "Verifiable Environments for AI in Biology — Kenny Workman, LatchBio",
    "url": "https://ai.engineer/talks/verifiable-environments-for-ai-in-biology-kenny-workman-latchbio",
    "videoId": "3ZMUiFaQ3qg",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1062000,
    "speakers": [
      {
        "slug": "kenny-workman",
        "name": "Kenny Workman",
        "organization": "LatchBio"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "LatchBio co-founder and CTO Kenny Workman explains why rapidly growing single-cell and spatial-biology datasets create opportunities for scientific AI agents. He presents SpatialBench, a benchmark of verifiable spatial-analysis tasks using deterministic graders, and describes how expert review uncovers ambiguous task specifications and problematic grading thresholds. He then introduces SpatialBench-Long for realistic, multistep biological investigations and argues that rigorous benchmarks can drive improvements in scientific AI models."
  },
  {
    "slug": "vibe-coding-at-scale-customizing-ai-assistants-for-enterprise-environments",
    "title": "Vibe Coding at Scale: Customizing AI Assistants for Enterprise Environments",
    "url": "https://ai.engineer/talks/vibe-coding-at-scale-customizing-ai-assistants-for-enterprise-environments",
    "videoId": "EAfP8pDs7h4",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 4838000,
    "speakers": [
      {
        "slug": "harald-kirschner",
        "name": "Harald Kirschner",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "An interactive enterprise-oriented vibe-coding workshop contrasts outcome-first experimentation with more disciplined AI-assisted development, demonstrates model-generated application design, and discusses agent auto-approval and associated safety concerns. The presenter explores external tools and reusable prompts before explaining Model Context Protocol integration, the transition from legacy HTTP-plus-SSE to Streamable HTTP, and server-initiated sampling through a client's language model."
  },
  {
    "slug": "vibe-coding-at-scale-customizing-ai-assistants-for-enterprise-environments-harald-kirschner",
    "title": "Vibe Coding at Scale: Customizing AI Assistants for Enterprise Environments - Harald Kirschner",
    "url": "https://ai.engineer/talks/vibe-coding-at-scale-customizing-ai-assistants-for-enterprise-environments-harald-kirschner",
    "videoId": "i1uPAN6uW4s",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 926000,
    "speakers": [
      {
        "slug": "harald-kirschner",
        "name": "Harald Kirschner",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Microsoft's Harald Kirschner contrasts experimental, auto-accepting YOLO vibe coding with structured AI-assisted development that adds enterprise guardrails and maintainability. Demonstrating VS Code Insiders and Copilot-related workflows, he explores workspace-scoped auto-approval, local voice input, rapid React/Vite/Material UI prototyping, automatic saving, and how MCP integrations, task runners, and linters help coding agents operate effectively in larger codebases."
  },
  {
    "slug": "vibe-coding-with-confidence",
    "title": "Vibe Coding, with Confidence",
    "url": "https://ai.engineer/talks/vibe-coding-with-confidence",
    "videoId": "n991Yxo1aOI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1255000,
    "speakers": [
      {
        "slug": "itamar-friedman",
        "name": "Itamar Friedman",
        "organization": "Qodo"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Qodo CEO and co-founder Itamar Friedman argues that enterprise vibe coding requires reliable, context-aware workflows spanning the software development lifecycle, rather than faster code generation alone. He introduces a newly launched command-line tool, discusses testing, review, bug fixing, and refactoring, and describes how MCP and A2A can support coordinated swarms of specialized agents."
  },
  {
    "slug": "vibe-engineering-effect-apps",
    "title": "Vibe Engineering Effect Apps",
    "url": "https://ai.engineer/talks/vibe-engineering-effect-apps",
    "videoId": "Wmp2Tku2PrI",
    "event": "AI Engineer Europe 2026",
    "durationMs": 6184000,
    "speakers": [
      {
        "slug": "michael-arnaldi",
        "name": "Michael Arnaldi",
        "organization": "Effectful Technologies Inc"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Effect creator Michael Arnaldi leads an interactive, from-scratch workshop on building reliable Effect applications with coding agents. He argues that agents work best when they can inspect the library’s upstream repository directly, then explores an empty Bun/Vitest project, agent guidance, HTTP API patterns, edge validation, and Effect layer practices while discussing LLM limitations and audience experience."
  },
  {
    "slug": "vibes-won-t-cut-it",
    "title": "Vibes won't cut it",
    "url": "https://ai.engineer/talks/vibes-won-t-cut-it",
    "videoId": "Dc3qOA9WOnE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 934000,
    "speakers": [
      {
        "slug": "chris-kelly",
        "name": "Chris Kelly",
        "organization": "Augment Code"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Augment Code's Chris Kelly argues that generating code is not equivalent to engineering production software. He explains why AI-generated code still requires architectural judgment, maintenance, debugging, operational accountability, and deep codebase context, while predicting that software engineers will remain essential and code review will become increasingly valuable."
  },
  {
    "slug": "video-has-no-memory-here-s-how-we-built-one",
    "title": "Video Has No Memory. Here's How We Built One.",
    "url": "https://ai.engineer/talks/video-has-no-memory-here-s-how-we-built-one",
    "videoId": "mOf-PP4mVjA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1227000,
    "speakers": [
      {
        "slug": "james-le",
        "name": "James Le",
        "organization": "TwelveLabs"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "James Le explains why video AI needs durable, configurable memory rather than repeatedly answering queries from scratch. He presents video as a multimodal spatiotemporal volume and describes TwelveLabs' stack: Marengo embeddings, Pegasus video-language reasoning, API-accessible context graphs, and the Jockey video-agent product. Demonstrations cover tracking Lionel Messi across footage, traffic-safety detection, and advertising use cases, followed by knowledge stores, corpus digests, and agentic search."
  },
  {
    "slug": "viktor-ai-coworker-that-lives-in-slack",
    "title": "Viktor — AI Coworker That Lives in Slack",
    "url": "https://ai.engineer/talks/viktor-ai-coworker-that-lives-in-slack",
    "videoId": "ohKt066uFhg",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1170000,
    "speakers": [
      {
        "slug": "fryderyk-wiatrowski",
        "name": "Fryderyk Wiatrowski",
        "organization": "Viktor"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Viktor co-founder Fryderyk Wiatrowski explains how the company evolved from browser and email agents into a Slack-native AI coworker with shared access to company tools. He focuses on the engineering difficulties of organization-wide agents: complex Slack interactions, persistent shared context, isolation between channels and direct messages, integration permissions, and preventing private email or executive-channel information from leaking across teams."
  },
  {
    "slug": "vision-ai-in-2025-peter-robicheaux-roboflow",
    "title": "Vision AI in 2025 — Peter Robicheaux, Roboflow",
    "url": "https://ai.engineer/talks/vision-ai-in-2025-peter-robicheaux-roboflow",
    "videoId": "IQc05eCvNYE",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1044000,
    "speakers": [
      {
        "slug": "peter-robicheaux",
        "name": "Peter Robicheaux",
        "organization": "Roboflow"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Roboflow ML lead Peter Robicheaux argues that practical computer vision demands low-latency edge deployment and stronger visual reasoning than saturated ImageNet and COCO benchmarks capture. He contrasts CLIP's vision-language representations with DINOv2's self-supervised visual features, examines multimodal perception failures and the different pretraining behavior of convolutional YOLOv8 versus transformer-based LW-DETR, and introduces RF100-VL, a curated benchmark spanning 100 real-world object-detection datasets. Audience questions address edge fine-tuning and Roboflow's support for researchers."
  },
  {
    "slug": "vision-zero-bugs",
    "title": "Vision: Zero Bugs",
    "url": "https://ai.engineer/talks/vision-zero-bugs",
    "videoId": "qLqttdO33UM",
    "event": "AI Engineer Code 2025",
    "durationMs": 2163000,
    "speakers": [
      {
        "slug": "johann-schleier-smith",
        "name": "Johann Schleier-Smith",
        "organization": "Temporal Technologies"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Temporal AI technical lead Johann Schleier-Smith argues that AI can make traditionally expensive, high-assurance software engineering practical beyond aerospace. Drawing on the Airbus A320 and Curiosity rover, he discusses N-version programming, defensive programming, static analysis, structured programming, modular decomposition, and independent verification as foundations for more reliable LLM-generated software, while positioning durable execution as complementary reliability infrastructure."
  },
  {
    "slug": "voice-agents-the-good-the-bad-and-the-ugly",
    "title": "Voice Agents: the good, the bad, and the ugly",
    "url": "https://ai.engineer/talks/voice-agents-the-good-the-bad-and-the-ugly",
    "videoId": "pzmbleiOfCM",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1128000,
    "speakers": [
      {
        "slug": "eddie-siegel",
        "name": "Eddie Siegel",
        "organization": "Fractional AI"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Fractional AI CTO Eddie Siegel explains how his team built a production voice agent for large-scale qualitative research interviews. He demonstrates the agent, describes using the OpenAI Realtime API alongside tool calls and background text-based agents to monitor interview progress and select follow-up questions, and explains how synthetic conversations and evaluations expose regressions and guide development despite subjective conversational outcomes."
  },
  {
    "slug": "voice-agents-with-realtime-video-sidney-primas-lemonslice",
    "title": "Voice agents with Realtime Video — Sidney Primas, LemonSlice",
    "url": "https://ai.engineer/talks/voice-agents-with-realtime-video-sidney-primas-lemonslice",
    "videoId": "z1dqv74SpUs",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1596000,
    "speakers": [
      {
        "slug": "sidney-primas",
        "name": "Sidney Primas",
        "organization": "LemonSlice"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      }
    ],
    "summary": "LemonSlice cofounder and CTO Sidney Primas presents real-time visual agents intended to approach an avatar Turing test. He demonstrates an interactive historical avatar and single-image characters, explains how LemonSlice adds an API-based visual layer to independently supplied language and voice models, and discusses model orchestration, emotional expression, evaluation, and future operating costs."
  },
  {
    "slug": "voice-ai-your-bot-isn-t-special",
    "title": "Voice AI: Your Bot Isn't Special",
    "url": "https://ai.engineer/talks/voice-ai-your-bot-isn-t-special",
    "videoId": "2p2ErKRELHM",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1147000,
    "speakers": [
      {
        "slug": "nik-caryotakis",
        "name": "Nik Caryotakis",
        "organization": "SuperDial"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "SuperDial engineer Nik Caryotakis explains how to build reliable production voice agents for healthcare insurance calls, emphasizing conversation design, phone-tree navigation, EHR integrations, and last-mile operational reliability over voice realism. He demonstrates an insurance-call workflow and describes a practical stack using Pipecat for orchestration, TensorZero for generative-response routing, and Deepgram for speech recognition, alongside processor-level latency monitoring and upgradeable models."
  },
  {
    "slug": "voice-in-visuals-out-the-agony-and-the-ecstasy",
    "title": "Voice In, Visuals Out: The Agony and the Ecstasy",
    "url": "https://ai.engineer/talks/voice-in-visuals-out-the-agony-and-the-ecstasy",
    "videoId": "65X0pQ6Lmbg",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 785000,
    "speakers": [
      {
        "slug": "allen-pike",
        "name": "Allen Pike",
        "organization": "Forestwalk Labs"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Forestwalk Labs co-founder Allen Pike explains how AI agents can combine spoken input with rich visual output and describes a meeting assistant that creates work items during conversation. He emphasizes the approximately 200-millisecond responsiveness needed for seamless interaction, contrasts observed GPT-5 mini and Haiku latency, and recommends latency-oriented inference and stable prefix caching."
  },
  {
    "slug": "voiceops-fying-low-latency-intelligence-extraction-from-messy-audio-streams-dippu-kumar-singh",
    "title": "VoiceOps-fying Low-Latency Intelligence Extraction from Messy Audio Streams — Dippu Kumar Singh",
    "url": "https://ai.engineer/talks/voiceops-fying-low-latency-intelligence-extraction-from-messy-audio-streams-dippu-kumar-singh",
    "videoId": "IEF842ZEU5A",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1376000,
    "speakers": [
      {
        "slug": "dippu-kumar-singh",
        "name": "Dippu Kumar Singh",
        "organization": "Fujitsu North America"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Fujitsu North America's Dippu Kumar Singh presents a low-latency contact-center voice-intelligence pipeline that captures multi-channel audio, applies speech-to-text, extracts structured customer intent and operator actions with prompted LLM workflows, and synchronizes verified results into CRM systems. He highlights after-call work reduction, channel separation, early PII masking, transcription accuracy, human-in-the-loop confirmation, and more consistent operational data."
  },
  {
    "slug": "voicevision-rag-integrating-visual-document-intelligence-with-voice-response",
    "title": "VoiceVision RAG - Integrating Visual Document Intelligence with Voice Response",
    "url": "https://ai.engineer/talks/voicevision-rag-integrating-visual-document-intelligence-with-voice-response",
    "videoId": "hwCmfThIiS4",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 5032000,
    "speakers": [
      {
        "slug": "suman-debnath",
        "name": "Suman Debnath",
        "organization": "Amazon Web Services (AWS)"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "AWS principal machine learning advocate Suman Debnath presents a hands-on workshop combining ColPali visual document retrieval with the open-source Strands Agents SDK and synthesized voice responses. Using a science textbook as an example, he demonstrates document-page retrieval, agent-directed summarization and tool use, then experiments with prompts that control spoken-answer delivery."
  },
  {
    "slug": "waymo-s-emma-teaching-cars-to-think-jyh-jing-hwang-waymo",
    "title": "Waymo's EMMA: Teaching Cars to Think - Jyh-Jing Hwang, Waymo",
    "url": "https://ai.engineer/talks/waymo-s-emma-teaching-cars-to-think-jyh-jing-hwang-waymo",
    "videoId": "iS9YFW28XyM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1048000,
    "speakers": [
      {
        "slug": "jyh-jing-hwang",
        "name": "Jyh-Jing Hwang",
        "organization": "Waymo"
      }
    ],
    "topics": [
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Waymo researcher Jyh-Jing Hwang explains how EMMA adapts Gemini into an end-to-end multimodal autonomous-driving model that reasons about unusual, safety-critical road situations and converts camera inputs and navigation instructions into planned trajectories. He discusses nuScenes and Waymo Open Motion Dataset results, unified driving and 3D detection tasks, scaling with additional training data, and evaluating driving models with realistic generated video including Veo 2 examples."
  },
  {
    "slug": "we-cut-94-of-our-ai-coding-tokens-with-a-local-code-index-here-s-the-architecture",
    "title": "We Cut 94% of Our AI Coding Tokens With a Local Code Index. Here's the Architecture.",
    "url": "https://ai.engineer/talks/we-cut-94-of-our-ai-coding-tokens-with-a-local-code-index-here-s-the-architecture",
    "videoId": "dRmWYHuIJxM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 643000,
    "speakers": [
      {
        "slug": "rajkumar-sakthivel",
        "name": "Rajkumar Sakthivel",
        "organization": "Tesco"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Rajkumar Sakthivel explains Code Context Engine, a local code-retrieval layer that reduces unnecessary context sent to AI coding tools through indexed snippets, cross-file relationships, lightweight relevance scoring, and shared project memory. He describes a FastAPI benchmark, a 0.4-millisecond heuristic that avoids extra LLM calls, and a reported 94% token reduction, while stressing that the percentage compares against a worst-case full-file baseline rather than smarter existing coding agents."
  },
  {
    "slug": "we-gave-an-agent-production-code-access-and-then-tried-to-sleep-at-night",
    "title": "We Gave an Agent Production Code Access and Then Tried to Sleep at Night",
    "url": "https://ai.engineer/talks/we-gave-an-agent-production-code-access-and-then-tried-to-sleep-at-night",
    "videoId": "LqLoYksJ6do",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1317000,
    "speakers": [
      {
        "slug": "moritz-johner",
        "name": "Moritz Johner",
        "organization": "Form3"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Form3's Moritz Johner presents PatchPilot, a production coding agent for remediating CVEs across thousands of repositories when ordinary dependency updaters cannot address vulnerabilities spanning base images, runtime versions, and CI failures. He argues that agents with production credentials become software supply-chain actors, and describes an architecture separating deterministic Go orchestration from constrained agentic remediation, with minimal changes, restricted network access, and human-reviewed pull requests."
  },
  {
    "slug": "we-vetted-2-000-ai-skills-before-they-reached-developers",
    "title": "We Vetted 2,000 AI Skills Before They Reached Developers",
    "url": "https://ai.engineer/talks/we-vetted-2-000-ai-skills-before-they-reached-developers",
    "videoId": "iKQ78wyJEXU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 984000,
    "speakers": [
      {
        "slug": "lucas-palma",
        "name": "Lucas Palma",
        "organization": "Nubank"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Nubank Product Security Manager Lucas Palma explains how the company vetted 2,000 AI skills before developer distribution. He frames skills, plugins, MCP servers, and agent rules as an expanded software supply chain, then describes Skill Vector: a local and CI-integrated review system combining deterministic security checks with contextual LLM analysis, pull-request feedback, and policy-based remediation or blocking. The workflow targets credential leaks, destructive shell commands, behavioral drift, overbroad permissions, exposed data, and risky MCP usage in a regulated financial-services environment."
  },
  {
    "slug": "wearing-the-agent-engineering-a-family-and-friends-personal-agent-from-group-chats-to-glasses",
    "title": "Wearing the Agent: Engineering a Family-and-Friends Personal Agent, from Group Chats to Glasses",
    "url": "https://ai.engineer/talks/wearing-the-agent-engineering-a-family-and-friends-personal-agent-from-group-chats-to-glasses",
    "videoId": "s67bE2Ur3bY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1149000,
    "speakers": [
      {
        "slug": "sai-krishna-rallabandi",
        "name": "Sai Krishna Rallabandi",
        "organization": "Fidelity Investments"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Sai Krishna Rallabandi of Fidelity Investments presents Judith, a personal AI agent deployed across family-and-friends group chats, and explains why shared agents require stronger privacy controls, role-aware routing, and durable memory than [REDACTED:marital_status]-user assistants. He describes the Jataayu security layer, compositional risks involving OCR and reporting tools, separation of instructions from untrusted data using LoRA-fine-tuned small language models, and approaches to memory curation, evaluation, KV-cache-aware serving, and determining when an agent should respond."
  },
  {
    "slug": "what-breaks-when-you-build-ai-under-sovereignty-constraints",
    "title": "What Breaks When You Build AI Under Sovereignty Constraints",
    "url": "https://ai.engineer/talks/what-breaks-when-you-build-ai-under-sovereignty-constraints",
    "videoId": "x2bH0RKPgdc",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1149000,
    "speakers": [
      {
        "slug": "bilge-yucel",
        "name": "Bilge Yücel",
        "organization": "deepset GmbH"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Bilge Yücel of deepset examines sovereign AI as an engineering problem requiring explicit control over data flows, model selection, infrastructure, and operations. She explains how cross-border processing, provider dependence, self-hosting, Kubernetes and GPU requirements, monitoring, and auditability complicate existing AI systems, then illustrates how Haystack supports modular sovereign agent architectures, guardrails, retrieval pipelines, and reusable tools."
  },
  {
    "slug": "what-data-from-20-million-pull-requests-reveal-about-ai-transformation",
    "title": "What Data from 20 Million Pull Requests Reveal About AI Transformation",
    "url": "https://ai.engineer/talks/what-data-from-20-million-pull-requests-reveal-about-ai-transformation",
    "videoId": "WqZq8L-v9pA",
    "event": "AI Engineer Code 2025",
    "durationMs": 1077000,
    "speakers": [
      {
        "slug": "nicholas-arcolano",
        "name": "Nicholas Arcolano",
        "organization": "Jellyfish"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Jellyfish head of research Nicholas Arcolano examines evidence from 20 million pull requests to explain how engineering organizations adopt AI coding tools and what productivity changes follow. He defines adoption through observed developer behavior, reports median usage increasing from approximately 22% to nearly 90%, and notes that fully autonomous agents account for less than 0.2% of merged pull requests. The talk connects AI adoption with pull-request throughput, cycle time, and 18% larger pull requests, while emphasizing that repository architecture and context limitations can constrain productivity gains."
  },
  {
    "slug": "what-do-models-still-suck-at",
    "title": "What Do Models Still Suck At?",
    "url": "https://ai.engineer/talks/what-do-models-still-suck-at",
    "videoId": "R7A8rX-09Zw",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1224000,
    "speakers": [
      {
        "slug": "peter-gostev",
        "name": "Peter Gostev",
        "organization": "Arena.ai"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Peter Gostev of Arena.ai argues that improving benchmark scores conceal persistent real-world language-model failures. He introduces BullshitBench, which uses nonsensical prompts and LLM-as-a-judge grading to test whether models challenge invalid premises, compares Claude, GPT, Gemini, Qwen, and Grok, and examines reasoning traces. He also presents Arena voting and dissatisfaction data across time and domains, including medicine, finance, law, and game development."
  },
  {
    "slug": "what-does-done-even-mean-agents-and-paperclip-s-liveness-model-dotta-paperclip",
    "title": "What Does Done Even Mean? Agents and Paperclip's Liveness Model - Dotta, Paperclip",
    "url": "https://ai.engineer/talks/what-does-done-even-mean-agents-and-paperclip-s-liveness-model-dotta-paperclip",
    "videoId": "7P0elyLIxXo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 434000,
    "speakers": [
      {
        "slug": "dotta",
        "name": "Dotta",
        "organization": "Paperclip"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Paperclip creator Dotta argues that agent-generated work cannot safely be considered complete based on a single status flag: merging, deploying, and announcing require different evidence, verification, authority, and risk ownership. He explains how Paperclip balances workflow liveness with quality control through explicit task-state transitions, harness-agnostic watchdogs, structured completion claims, and independent cross-model review."
  },
  {
    "slug": "what-does-enterprise-ready-mcp-mean",
    "title": "What does Enterprise Ready MCP mean?",
    "url": "https://ai.engineer/talks/what-does-enterprise-ready-mcp-mean",
    "videoId": "0MqYA52iWQU",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 833000,
    "speakers": [
      {
        "slug": "tobin-south",
        "name": "Tobin South",
        "organization": "WorkOS"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "Tobin South, representing WorkOS at the time of the presentation, explains what enterprise readiness requires for Model Context Protocol servers and AI-agent workloads: standardized tool connections, authentication and authorization, provisioning, oversight, abuse prevention and data loss prevention. A hypothetical goat-feeding application illustrates scaling and payments, followed by discussion of Block’s goose and a live MCP-enabled merchandise purchase."
  },
  {
    "slug": "what-does-it-take-to-build-a-personal-local-private-ai-agent-that-augments-you-deeply",
    "title": "What does it take to build a personal, local, private AI Agent that augments you deeply?",
    "url": "https://ai.engineer/talks/what-does-it-take-to-build-a-personal-local-private-ai-agent-that-augments-you-deeply",
    "videoId": "jMoAaZP_Kkw",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1232000,
    "speakers": [
      {
        "slug": "soumith-chintala",
        "name": "Soumith Chintala",
        "organization": "Meta PyTorch"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "PyTorch co-founder Soumith Chintala examines what personal AI agents need to act reliably while keeping sensitive context under their users' control. He contrasts agents with news aggregators, shows how fragmented access to email, messaging, and financial services produces misleading answers, and discusses battery and mobile-platform limitations. He recommends an always-on Mac mini as a practical local host and highlights privacy, delegated-action risks, cloud-provider trust, and the need for better catastrophic-action classifiers."
  },
  {
    "slug": "what-every-ai-engineer-needs-to-know-about-gpus",
    "title": "What every AI engineer needs to know about GPUs",
    "url": "https://ai.engineer/talks/what-every-ai-engineer-needs-to-know-about-gpus",
    "videoId": "y-UGrYbJsJk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1192000,
    "speakers": [
      {
        "slug": "charles-frye",
        "name": "Charles Frye",
        "organization": "Modal"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Charles Frye of Modal explains why AI application engineers need practical GPU literacy as open-weight models and self-hosted inference become more accessible. He emphasizes tensor cores, low-precision matrix multiplication, bandwidth over latency, and GPU parallelism; discusses Dynamo, vLLM, SGLang, and TensorRT-LLM; and connects these hardware constraints to decoding strategies and serverless Python inference infrastructure."
  },
  {
    "slug": "what-if-the-harness-mattered-more-than-the-model-aditya-bhargava-etsy",
    "title": "What if the harness mattered more than the model? - Aditya Bhargava, Etsy",
    "url": "https://ai.engineer/talks/what-if-the-harness-mattered-more-than-the-model-aditya-bhargava-etsy",
    "videoId": "2e9ANoOEn28",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1924000,
    "speakers": [
      {
        "slug": "aditya-bhargava",
        "name": "Aditya Bhargava",
        "organization": "Etsy"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Etsy staff engineer Aditya Bhargava argues that agent performance, especially with weaker models, depends heavily on the surrounding harness rather than model capability alone. He demonstrates Agency, his TypeScript-inspired agent programming language, progressing from basic LLM calls to automatically generated tools and JSON Schema, restricted filesystem access through partial function application, ReAct-style test-and-repair feedback loops, and context-efficient subagents."
  },
  {
    "slug": "what-if-the-network-was-the-sandbox",
    "title": "What if the network was the sandbox?",
    "url": "https://ai.engineer/talks/what-if-the-network-was-the-sandbox",
    "videoId": "BM2JX9hqsVQ",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1469000,
    "speakers": [
      {
        "slug": "remy-guercio",
        "name": "Remy Guercio",
        "organization": "Tailscale"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Remy Guercio argues that sandboxing AI agents requires separating execution isolation from credential and access control. He explains how Tailscale layers user, group, and workload identity over WireGuard connections, then demonstrates Aperture, an identity-aware AI gateway that centralizes provider credentials and exposes request-level usage and authorization for developers and automated agents."
  },
  {
    "slug": "what-is-a-humanoid-foundation-model-an-introduction-to-gr00t-n1",
    "title": "What Is a Humanoid Foundation Model? An Introduction to GR00T N1",
    "url": "https://ai.engineer/talks/what-is-a-humanoid-foundation-model-an-introduction-to-gr00t-n1",
    "videoId": "mWKYvT9Lc50",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1067000,
    "speakers": [
      {
        "slug": "annika-brundyn",
        "name": "Annika Brundyn",
        "organization": "NVIDIA"
      },
      {
        "slug": "aastha-jhunjhunwala",
        "name": "Aastha Jhunjhunwala",
        "organization": "NVIDIA"
      }
    ],
    "topics": [
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "NVIDIA speakers Annika Brundyn and Aastha Jhunjhunwala introduce GR00T N1, a humanoid vision-language-action foundation model. They explain physical-AI labor applications; the simulation, training, and edge-deployment lifecycle across OVX, DGX, and AGX; scarce robotics demonstrations and synthetic-data generation; and the model’s cognition-inspired architecture, imitation learning, and manipulation examples."
  },
  {
    "slug": "what-it-actually-takes-to-deploy-genai-applications-to-enterprises",
    "title": "What It Actually Takes to Deploy GenAI Applications to Enterprises",
    "url": "https://ai.engineer/talks/what-it-actually-takes-to-deploy-genai-applications-to-enterprises",
    "videoId": "42q8OmAF_Gw",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1290000,
    "speakers": [
      {
        "slug": "arjun-bansal",
        "name": "Arjun Bansal",
        "organization": "Log10"
      },
      {
        "slug": "trey-doig",
        "name": "Trey Doig",
        "organization": "Echo AI"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Arjun Bansal of Log10 and Trey Doig of Echo AI explain how enterprise conversation-intelligence systems ingest and normalize customer interactions, apply configurable LLM pipelines, and surface insights across entire conversation datasets. They discuss earning customer trust through accurate evaluations, using Log10 AutoFeedback to address LLM-as-a-judge weaknesses and track hallucinations, and supporting production-scale throughput with self-hosted, domain-adapted models."
  },
  {
    "slug": "what-lies-beneath-the-api-benjamin-cowen-modal",
    "title": "What Lies Beneath the API — Benjamin Cowen, Modal",
    "url": "https://ai.engineer/talks/what-lies-beneath-the-api-benjamin-cowen-modal",
    "videoId": "HvZXAOZ3iv8",
    "event": "AI Engineer Europe 2026",
    "durationMs": 760000,
    "speakers": [
      {
        "slug": "benjamin-cowen",
        "name": "Benjamin Cowen",
        "organization": "Modal"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Modal engineer Benjamin Cowen argues that mature, specialized AI applications can outgrow general-purpose frontier APIs and benefit from fine-tuned models when inference economics or application-specific performance become limiting. He describes serverless infrastructure for training, sandboxed reinforcement-learning rollouts, and autoscaling custom inference with tools including vLLM, SGLang, and NVIDIA Triton Inference Server."
  },
  {
    "slug": "what-rl-means-for-agents",
    "title": "What RL Means for Agents",
    "url": "https://ai.engineer/talks/what-rl-means-for-agents",
    "videoId": "JIsgyk0Paic",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1097000,
    "speakers": [
      {
        "slug": "will-brown",
        "name": "Will Brown",
        "organization": "Morgan Stanley"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Morgan Stanley machine learning researcher Will Brown examines how reinforcement learning could move language-model systems beyond engineered workflows toward more autonomous agents. He discusses reasoning models and test-time scaling, DeepSeek-R1 and GRPO, emergent chain-of-thought behavior, OpenAI deep research as an example of end-to-end reinforcement learning for tool-using agents, and experimental reward shaping for agent research."
  },
  {
    "slug": "what-s-new-from-anthropic-and-what-s-next",
    "title": "What's new from Anthropic and what's next",
    "url": "https://ai.engineer/talks/what-s-new-from-anthropic-and-what-s-next",
    "videoId": "EuC1GWhQdKE",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 818000,
    "speakers": [
      {
        "slug": "alex-albert",
        "name": "Alex Albert",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Anthropic's Alex Albert argues that useful AI products must be redesigned around LLM capabilities rather than adding superficial AI features. He introduces Claude 3.5 Sonnet, compares it with Claude 3 Opus, and discusses evaluation benchmarks, a 200K-token context window, retrieval, coding, OCR, and availability through the Anthropic API, Amazon Bedrock, and Vertex AI. He presents Artifacts and Projects as examples of AI-native collaboration, then describes the Tool Use API, custom client-side functions, and structured JSON output."
  },
  {
    "slug": "what-s-next-after-rlhf",
    "title": "What's next after RLHF?",
    "url": "https://ai.engineer/talks/what-s-next-after-rlhf",
    "videoId": "cJ0EOzey--o",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1085000,
    "speakers": [
      {
        "slug": "diogo-almeida",
        "name": "Diogo Almeida",
        "organization": "TypeSafe AI"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "TypeSafe AI CEO Diogo Almeida, an InstructGPT and GPT-4 paper coauthor, argues that RLHF optimized language models for pleasing humans and interactive assistance rather than reliable autonomous decisions. He contrasts impressive benchmarks and chatbot-like products with limited practical automation, discusses Claude Code, RLVR, and reward-model asymmetry, and calls for automation-native software that can perform consequential work without continuous human oversight."
  },
  {
    "slug": "what-the-best-agents-share",
    "title": "What the Best Agents Share",
    "url": "https://ai.engineer/talks/what-the-best-agents-share",
    "videoId": "7CrPrHgoEYk",
    "event": "AI Engineer Europe 2026",
    "durationMs": 621000,
    "speakers": [
      {
        "slug": "mardu-swanepoel",
        "name": "Mardu Swanepoel",
        "organization": "Flinn AI"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Mardu Swanepoel of Flinn AI examines four design patterns shared by effective AI agents: task-specific modes that narrow the action space and clarify expectations; transparent execution that makes progress visible and enables intervention; personalization and memory that increase speed to understanding; and reversible actions that bound mistakes. Examples include Cursor, Claude Cowork, Manus, Harvey, Notion, and Microsoft Word integration."
  },
  {
    "slug": "what-we-learned-deploying-ai-within-bloomberg-s-engineering-organization",
    "title": "What We Learned Deploying AI within Bloomberg’s Engineering Organization",
    "url": "https://ai.engineer/talks/what-we-learned-deploying-ai-within-bloomberg-s-engineering-organization",
    "videoId": "Q81AzlA-VE8",
    "event": "AI Engineer Code 2025",
    "durationMs": 1101000,
    "speakers": [
      {
        "slug": "lei-zhang",
        "name": "Lei Zhang",
        "organization": "Bloomberg"
      }
    ],
    "topics": [
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Bloomberg infrastructure engineering leader Lei Zhang describes deploying AI across an organization of more than 9,000 engineers. Initial coding-assistant gains in prototyping, tests, and scripts did not resolve broader engineering bottlenecks, motivating agents for maintenance, migration, refactoring, and incident response. He discusses deterministic verification, growing pull-request and merge queues, MCP connections to operational telemetry and service topology, open-source contributions to KServe and Envoy AI Gateway, and the importance of reliability, training, and engineering leadership."
  },
  {
    "slug": "what-we-learned-from-a-year-of-building-with-llms",
    "title": "What We Learned From A Year of Building With LLMs",
    "url": "https://ai.engineer/talks/what-we-learned-from-a-year-of-building-with-llms",
    "videoId": "qBHfQT3YtyY",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 2121000,
    "speakers": [
      {
        "slug": "eugene-yan",
        "name": "Eugene Yan",
        "organization": "Amazon"
      },
      {
        "slug": "hamel-husain",
        "name": "Hamel Husain",
        "organization": "Parlance Labs"
      },
      {
        "slug": "jason-liu",
        "name": "Jason Liu"
      },
      {
        "slug": "dr-bryan-bischof",
        "name": "Dr Bryan Bischof",
        "organization": "Hex"
      },
      {
        "slug": "charles-frye",
        "name": "Charles Frye",
        "organization": "Modal"
      },
      {
        "slug": "shreya-shankar",
        "name": "Shreya Shankar",
        "organization": "UC Berkeley"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Six coauthors translate a year of real-world LLM application experience into strategic, operational, and tactical guidance. They argue that models themselves are rarely a durable moat, discuss switching providers as capabilities and costs change, examine AI engineering roles and team workflows, and emphasize production evaluations, monitoring, guardrails, and the time required to move beyond demos."
  },
  {
    "slug": "what-we-learned-from-using-llms-in-pinterest",
    "title": "What We Learned from Using LLMs in Pinterest",
    "url": "https://ai.engineer/talks/what-we-learned-from-using-llms-in-pinterest",
    "videoId": "XdAWgO11zuk",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1093000,
    "speakers": [
      {
        "slug": "mukuntha-narayanan",
        "name": "Mukuntha Narayanan",
        "organization": "Pinterest"
      },
      {
        "slug": "han-wang",
        "name": "Han Wang",
        "organization": "Pinterest"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      }
    ],
    "summary": "Pinterest Search engineers Han Wang and Mukuntha Narayanan explain how they improve search relevance using fine-tuned LLM cross-encoders that classify query-Pin relevance, enriched by vision-language-model captions and user signals. An 8-billion-parameter Llama 3 model is reported to outperform multilingual BERT by 12% and SearchSage by 20%. To serve the approach efficiently at scale, they distill a multilingual teacher into a production student model, selectively refresh embeddings, and reuse relevance-tuned representations across languages and downstream tasks. Audience questions cover model selection, ranking integration, earlier search architectures, multimodality, and multilingual results."
  },
  {
    "slug": "when-agents-meet-physical-data-the-other-physics-of-agent-harnesses",
    "title": "When Agents Meet Physical Data: The Other Physics of Agent Harnesses",
    "url": "https://ai.engineer/talks/when-agents-meet-physical-data-the-other-physics-of-agent-harnesses",
    "videoId": "bUJgirn4_yc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1653000,
    "speakers": [
      {
        "slug": "dmitry-petrov",
        "name": "Dmitry Petrov",
        "organization": "DataChain"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "DataChain co-founder Dmitry Petrov explains why coding agents struggle with physical-world and unstructured datasets: repeated perception, scattered object-storage metadata, and expensive recomputation make ordinary agent loops impractical. He demonstrates an open-source DataChain harness that converts dashcam-video analysis into persistent, queryable datasets using YOLO, Pydantic schemas, Python/SQL workflows, and execution infrastructure, enabling subsequent questions to reuse materialized results rather than rescan raw files."
  },
  {
    "slug": "when-all-context-matters-extended-cache-augmented-generation-ecag",
    "title": "When all context matters: Extended Cache Augmented Generation (ECAG)",
    "url": "https://ai.engineer/talks/when-all-context-matters-extended-cache-augmented-generation-ecag",
    "videoId": "XovaGv4f39A",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 352000,
    "speakers": [
      {
        "slug": "luis-romero-sevilla",
        "name": "Luis Romero-Sevilla",
        "organization": "Orbis Operations"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Orbis Operations VP of AI Luis Romero-Sevilla presents Extended Cache Augmented Generation (ECAG) for rapidly changing document collections in which relevant information is distributed across densely interconnected records. He contrasts vector-based RAG and GraphRAG with cache-augmented generation, then proposes parallel cached context buckets interrogated by a supervising model that synthesizes answers. He describes document-distribution challenges, claimed speed and accuracy advantages, and KV-cache lifetime, compute, and cost trade-offs."
  },
  {
    "slug": "when-vectors-break-down-graph-based-rag-for-dense-enterprise-knowledge",
    "title": "When Vectors Break Down: Graph-Based RAG for Dense Enterprise Knowledge",
    "url": "https://ai.engineer/talks/when-vectors-break-down-graph-based-rag-for-dense-enterprise-knowledge",
    "videoId": "XlAIgmi_Vow",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 947000,
    "speakers": [
      {
        "slug": "sam-julien",
        "name": "Sam Julien",
        "organization": "Writer"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Writer developer-relations director Sam Julien explains why vector-only retrieval breaks down when enterprise documents reuse similar terminology, illustrating how naive chunking and nearest-neighbor search can return incorrect facts. He traces Writer's progression from conventional search and vector embeddings to graph-based retrieval, discusses graph-database scaling and Cypher limitations, and describes storing graph-derived data as JSON in a Lucene-based search engine before combining knowledge graphs with Fusion-in-Decoder. He also situates the architecture alongside Writer's Palmyra enterprise models and domain-specific evaluation work."
  },
  {
    "slug": "when-will-the-benchmaxxing-plague-end",
    "title": "When Will The Benchmaxxing Plague End?",
    "url": "https://ai.engineer/talks/when-will-the-benchmaxxing-plague-end",
    "videoId": "-npY6XjM8CQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1045000,
    "speakers": [
      {
        "slug": "nick-heiner",
        "name": "Nick Heiner",
        "organization": "Surge AI"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Nick Heiner of Surge AI argues that benchmark optimization often diverges from real-world AI usefulness because evaluations are expensive to construct, vulnerable to leaderboard gaming and public-data contamination, and frequently rely on incomplete or exploitable verifiers. Examples include LMArena incentives, memorization of SWE-bench Verified, contradictory IFEval instructions, and Unicode-based reward hacking. He advocates higher-quality expert data, adversarially robust evaluation design, and complete two-way alignment between prompts and verifiers."
  },
  {
    "slug": "where-ai-is-superhuman-the-right-jobs-to-automate-with-llms",
    "title": "Where AI is superhuman: The right jobs to automate with LLMs",
    "url": "https://ai.engineer/talks/where-ai-is-superhuman-the-right-jobs-to-automate-with-llms",
    "videoId": "R2VfIjuWhgw",
    "event": "AI Engineer Summit 2025",
    "durationMs": 713000,
    "speakers": [
      {
        "slug": "andy-triedman",
        "name": "Andy Triedman",
        "organization": "Theory Ventures"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Theory Ventures partner Andy Triedman presents a framework for selecting enterprise jobs suited to LLM automation. He identifies transformation, synthesis, and reasoning as core capabilities, explains how task volume and complexity determine whether systems function as copilots or automate workflows, and discusses domain-specific data, human review, and escalation. A Dropzone AI security-operations example illustrates automated alert investigation, while later examples address personalization and evolving human responsibilities."
  },
  {
    "slug": "which-jobs-can-be-replaced-today",
    "title": "Which Jobs Can Be Replaced Today",
    "url": "https://ai.engineer/talks/which-jobs-can-be-replaced-today",
    "videoId": "R3hTescenDw",
    "event": "AI Engineer World's Fair 2024",
    "durationMs": 1199000,
    "speakers": [
      {
        "slug": "fryderyk-wiatrowski",
        "name": "Fryderyk Wiatrowski",
        "organization": "Zeta Labs"
      },
      {
        "slug": "peter-albert",
        "name": "Peter Albert",
        "organization": "Zeta Labs"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Zeta Labs cofounders Fryderyk Wiatrowski and Peter Albert explain how autonomous agents can absorb repetitive, low-leverage work while humans retain high-value decisions. They demonstrate Jace scheduling meetings through email, contrast reactive browser-based agents with limited APIs, and discuss practical agent-building techniques including familiar prompt formats, matching model training distributions, explicit state tracking, parallelization, execution feedback, model-based judging, and the costs of reinforcement learning."
  },
  {
    "slug": "while-my-guitar-gently-speaks",
    "title": "While my guitar gently speaks",
    "url": "https://ai.engineer/talks/while-my-guitar-gently-speaks",
    "videoId": "E_Txocq-Lrw",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1115000,
    "speakers": [
      {
        "slug": "todd-fisher",
        "name": "Todd Fisher",
        "organization": "Philo Ventures"
      }
    ],
    "topics": [
      {
        "slug": "creative-and-generative-media",
        "name": "Creative and generative media"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Todd Fisher of Philo Ventures demonstrates an AI-assisted guitar that speaks by combining audio software, pitch detection, synthesized notes, and generated speech. He introduces JUCE and digital audio workstation plug-ins, describes using the YIN pitch algorithm and ADSR synthesis, performs live demonstrations, and notes that computationally expensive processing must sometimes be prepared in advance. The presentation encourages engineers to pursue creative personal projects with AI."
  },
  {
    "slug": "why-agent-engineering",
    "title": "Why Agent Engineering",
    "url": "https://ai.engineer/talks/why-agent-engineering",
    "videoId": "5N33E9tC400",
    "event": "AI Engineer Summit 2025",
    "durationMs": 705000,
    "speakers": [
      {
        "slug": "swyx",
        "name": "swyx (Shawn Wang)",
        "organization": "Latent.Space"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Opening AI Engineer Summit’s agent-engineering day, swyx argues that AI engineering is emerging as a discipline distinct from machine learning and conventional software engineering. He explains the conference’s focus on agents, contrasts industry enthusiasm with earlier skepticism, discusses combining agents with retrieval, code generation, and search, and highlights improved reasoning, tool use, Model Context Protocol, and promising coding- and support-agent applications."
  },
  {
    "slug": "why-agent-hype-can-fall-short-of-reality-joel-becker-metr",
    "title": "Why Agent Hype can fall short of reality – Joel Becker, METR",
    "url": "https://ai.engineer/talks/why-agent-hype-can-fall-short-of-reality-joel-becker-metr",
    "videoId": "RhfqQKe22ZA",
    "event": "AI Engineer Code 2025",
    "durationMs": 1282000,
    "speakers": [
      {
        "slug": "joel-becker",
        "name": "Joel Becker",
        "organization": "METR"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "METR researcher Joel Becker contrasts rising AI benchmark and human-calibrated task-horizon results with a randomized field study of 16 experienced open-source developers. He explains why benchmark scores and autonomous task completion do not necessarily predict productivity on messy, context-dependent software work, while emphasizing uncertainty and the limitations of small studies."
  },
  {
    "slug": "why-and-how-you-need-to-sandbox-ai-generated-code-harshil-agrawal-cloudflare",
    "title": "Why, and how you need to sandbox AI-Generated Code? — Harshil Agrawal, Cloudflare",
    "url": "https://ai.engineer/talks/why-and-how-you-need-to-sandbox-ai-generated-code-harshil-agrawal-cloudflare",
    "videoId": "AHtGAgQ0Q_Q",
    "event": "AI Engineer Europe 2026",
    "durationMs": 2307000,
    "speakers": [
      {
        "slug": "harshil-agrawal",
        "name": "Harshil Agrawal",
        "organization": "Cloudflare"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Cloudflare developer educator Harshil Agrawal explains why AI-generated code must be treated as untrusted executable input, examining hallucinations, accidental credential exposure, and indirect prompt injection. He compares restricted V8 isolates with fuller Linux-container sandboxes, demonstrates capability-controlled Cloudflare Dynamic Workers, and uses PromptMotion to illustrate workloads requiring files, dependencies, background development servers, and live previews."
  },
  {
    "slug": "why-bolt-new-won-and-most-devtools-ai-pivots-failed-victoria-melnikova",
    "title": "Why Bolt.new Won and Most DevTools AI Pivots Failed - Victoria Melnikova",
    "url": "https://ai.engineer/talks/why-bolt-new-won-and-most-devtools-ai-pivots-failed-victoria-melnikova",
    "videoId": "3YRrBFeQ1aw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 345000,
    "speakers": [
      {
        "slug": "victoria-melnikova",
        "name": "Victoria Melnikova",
        "organization": "Evil Martians"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Evil Martians' Victoria Melnikova uses StackBlitz and Bolt.new to explain why superficial chatbot additions, overly broad AI ambitions, and waiting for perfect models undermine developer-tool pivots. Her three-step alternative is to identify an existing competitive advantage, use AI to amplify it, and redesign the workflow into a new product category; for Bolt.new, StackBlitz's WebContainers make prompt-driven, browser-based application development possible."
  },
  {
    "slug": "why-building-eval-platforms-is-hard",
    "title": "Why building eval platforms is hard",
    "url": "https://ai.engineer/talks/why-building-eval-platforms-is-hard",
    "videoId": "_fQ7Z_Wfouk",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1539000,
    "speakers": [
      {
        "slug": "phil-hetzel",
        "name": "Phil Hetzel",
        "organization": "Braintrust"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Braintrust solutions-engineering leader Phil Hetzel explains why reliable AI-agent evaluation requires more than running test inputs or recording results in spreadsheets. He connects LLM variability, cross-functional experimentation, prompt-configuration comparisons, production observability, and offline evaluation feedback loops, arguing that durable agent-quality platforms are fundamentally data and systems-engineering problems. Audience questions address dynamic prompts, Langfuse, and whether Braintrust must manage prompts directly."
  },
  {
    "slug": "why-can-t-anyone-answer-questions-about-the-business",
    "title": "Why Can't Anyone Answer Questions About the Business?",
    "url": "https://ai.engineer/talks/why-can-t-anyone-answer-questions-about-the-business",
    "videoId": "iUWwcG-C8OU",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1146000,
    "speakers": [
      {
        "slug": "garrett-galow",
        "name": "Garrett Galow",
        "organization": "WorkOS"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "WorkOS product leader Garrett Galow presents Studio, an internal AI workspace that lets employees answer business questions and create reusable, data-connected widgets without repeatedly asking engineers for SQL help. He describes a LangGraph agent using Opus, executable APIs and queries, grounding in current primary documentation, validation of Snowflake query results, and schema-aware operation without a dedicated RAG database. The presentation concludes with audience questions."
  },
  {
    "slug": "why-chatgpt-keeps-interrupting-you",
    "title": "Why ChatGPT Keeps Interrupting You",
    "url": "https://ai.engineer/talks/why-chatgpt-keeps-interrupting-you",
    "videoId": "1v9zBiZKlIY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1623000,
    "speakers": [
      {
        "slug": "tom-shapland-phd",
        "name": "Tom Shapland, PhD",
        "organization": "Livekit"
      }
    ],
    "topics": [
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "LiveKit product manager Tom Shapland explains why voice agents interrupt users: conventional cascaded pipelines often mistake silence-based voice activity detection for a completed conversational turn. Drawing on human turn-taking and parallel comprehension and response planning, he describes full-duplex approaches and demonstrates a LiveKit semantic model that considers the previous four conversational turns to reduce premature interruptions. Audience questions address visual cues, call and regeneration costs, benchmarking, and conversational backchannels."
  },
  {
    "slug": "why-eval-is-the-next-great-compute-primitive-sunil-pai-matt-carey-cloudflare",
    "title": "Why Eval++ Is the Next Great Compute Primitive — Sunil Pai & Matt Carey, Cloudflare",
    "url": "https://ai.engineer/talks/why-eval-is-the-next-great-compute-primitive-sunil-pai-matt-carey-cloudflare",
    "videoId": "SKDJo2CopRs",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1491000,
    "speakers": [
      {
        "slug": "sunil-pai",
        "name": "Sunil Pai",
        "organization": "Cloudflare"
      },
      {
        "slug": "matt-carey",
        "name": "Matt Carey",
        "organization": "Cloudflare"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Cloudflare engineers Sunil Pai and Matt Carey explain how Durable Objects provide addressable, persistent, low-latency state for collaborative AI agents. They discuss executing model-generated code inside capability-restricted Dynamic Workers, using Code Mode to expose Cloudflare's API through compact MCP tools, and supporting agent filesystems with SQLite and R2. Audience questions cover JavaScript, Python, WebAssembly, Zig, and the Vite-based vinext implementation of Next.js."
  },
  {
    "slug": "why-large-tiny-lms-agents-on-edge-robotics",
    "title": "Why Large? Tiny LMs & Agents on Edge/Robotics",
    "url": "https://ai.engineer/talks/why-large-tiny-lms-agents-on-edge-robotics",
    "videoId": "hacEQHHhu2Q",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1305000,
    "speakers": [
      {
        "slug": "cormac-brick",
        "name": "Cormac Brick",
        "organization": "Google"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "Google AI Edge technical lead Cormac Brick explains why deploying useful AI across inexpensive devices and robots requires smaller language models, especially as DRAM availability and cost constrain on-device inference. He surveys Google's LiteRT-LM, LiteRT, MediaPipe, and Gemma ecosystem, discusses quantization and Raspberry Pi deployment tradeoffs, and describes fine-tuning small models with synthetic data to support reliable voice-driven interactions at scale. A second, unnamed event facilitator briefly closes the session and announces that there is no time for audience Q&A."
  },
  {
    "slug": "why-mcp-and-chatgpt-apps-use-double-iframes-frederic-barthelet-alpic",
    "title": "Why MCP and ChatGPT Apps Use Double Iframes — Frédéric Barthelet, Alpic",
    "url": "https://ai.engineer/talks/why-mcp-and-chatgpt-apps-use-double-iframes-frederic-barthelet-alpic",
    "videoId": "c-2eEv2ou7Y",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1211000,
    "speakers": [
      {
        "slug": "frederic-barthelet",
        "name": "Frédéric Barthelet",
        "organization": "Alpic"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Alpic co-founder and CTO Frédéric Barthelet examines why MCP and ChatGPT applications require isolated iframe-based rendering for interactive tool-call views. He explains how Content Security Policy, script nonces, srcdoc, and browser storage complicate embedding third-party HTML, and highlights the risks of weakening host security. The talk concludes by pointing developers to Alpic's open-source Skybridge framework for building MCP and ChatGPT applications."
  },
  {
    "slug": "why-mlx",
    "title": "Why MLX",
    "url": "https://ai.engineer/talks/why-mlx",
    "videoId": "zTLJNHj0DeQ",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1390000,
    "speakers": [],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "[REDACTED:username] presents MLX as a foundation for running AI agents and multimodal models directly on Apple silicon instead of relying on cloud services. He explains MLX-VLM for accessible visual understanding, demonstrates local vision and Gemma-family model workflows, and describes MLX-Audio pipelines for low-latency speech synthesis, transcription, and voice interfaces. Community examples and audience questions illustrate practical applications and GPU-performance monitoring."
  },
  {
    "slug": "why-more-context-makes-your-agent-dumber-and-what-to-do-about-it",
    "title": "Why More Context Makes Your Agent Dumber and What to Do About It",
    "url": "https://ai.engineer/talks/why-more-context-makes-your-agent-dumber-and-what-to-do-about-it",
    "videoId": "EcqMYoIV57A",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1587000,
    "speakers": [
      {
        "slug": "nupur-sharma",
        "name": "Nupur Sharma",
        "organization": "Qodo"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Qodo's Nupur Sharma explains why expanding an agent's context window does not ensure that relevant information in the middle of its input is retained. She compares dependency mapping, knowledge graphs, and critic-based self-correction, emphasizing processing and latency costs, then describes an orchestration paradox in which capable models spend tokens planning rather than completing work. Qodo's code-review architecture collects pull-request context, coordinates specialized agents, and uses LangChain-based infrastructure; audience questions address agent communication and calibration."
  },
  {
    "slug": "why-off-the-shelf-ai-doesn-t-understand-money",
    "title": "Why Off-the-Shelf AI Doesn't Understand Money",
    "url": "https://ai.engineer/talks/why-off-the-shelf-ai-doesn-t-understand-money",
    "videoId": "Owb8g3yDyzo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1190000,
    "speakers": [
      {
        "slug": "udi-menkes",
        "name": "Udi Menkes",
        "organization": "Intuit"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Intuit principal product manager Udi Menkes argues that general-purpose LLMs produce a confident but unreliable financial-advice “fluent bluff” because they lack evidence about what happened when comparable businesses took specific actions. He describes grounding financial advisory systems in Intuit product data, causal treatment-effect measurement, millions of business trajectories, and reinforcement-learning models trained on state, action, and outcome relationships."
  },
  {
    "slug": "why-rust-is-the-ideal-language-for-vibe-coding",
    "title": "Why Rust is the Ideal Language for Vibe-Coding",
    "url": "https://ai.engineer/talks/why-rust-is-the-ideal-language-for-vibe-coding",
    "videoId": "ugUeZ8-b-u0",
    "event": "AI Engineer Europe 2026",
    "durationMs": 985000,
    "speakers": [
      {
        "slug": "daniel-szoke",
        "name": "Daniel Szoke",
        "organization": "Sentry"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Sentry Rust SDK maintainer Daniel Szoke argues that AI coding agents should optimize for correctness rather than merely producing runnable Python, JavaScript, or TypeScript quickly. Drawing on GitHub's TypeScript adoption data, limitations of permissive typing and testing, and a multithreaded counter example, he presents Rust's compile-time rejection of data races as a deterministic feedback mechanism that helps agents repair unsafe code before production. He closes by mentioning Sentry's agent-monitoring features."
  },
  {
    "slug": "why-senior-engineers-struggle-to-build-ai-agents",
    "title": "Why (Senior) Engineers Struggle to Build AI Agents",
    "url": "https://ai.engineer/talks/why-senior-engineers-struggle-to-build-ai-agents",
    "videoId": "3_gYbhABcAE",
    "event": "AI Engineer Europe 2026",
    "durationMs": 640000,
    "speakers": [
      {
        "slug": "philipp-schmid",
        "name": "Philipp Schmid",
        "organization": "Google DeepMind"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Google DeepMind’s Philipp Schmid explains why experienced software engineers can struggle with AI agents: agent development depends on natural-language context and semantic interpretation rather than strictly deterministic state. He discusses preserving execution by treating errors as inputs, evaluating outputs with traces, human experts, or LLM judges, and designing self-documenting tools and APIs that expose context developers otherwise leave implicit."
  },
  {
    "slug": "why-should-anyone-care-about-evals",
    "title": "Why should anyone care about Evals?",
    "url": "https://ai.engineer/talks/why-should-anyone-care-about-evals",
    "videoId": "jJ45Yz1lJao",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 341000,
    "speakers": [
      {
        "slug": "manu-goyal",
        "name": "Manu Goyal",
        "organization": "Braintrust"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Braintrust founding engineer Manu Goyal introduces the conference’s Evals track by explaining why stronger model metrics alone cannot justify deploying AI systems. Drawing on self-driving-car experience, he presents evaluations as a safe, fast experimentation loop rather than merely regression tests, then describes how Braintrust connects evaluations with prompt experimentation, logging, observability, and a data flywheel."
  },
  {
    "slug": "why-the-best-ai-agents-are-built-without-frameworks-primitives-over-frameworks",
    "title": "Why the Best AI Agents Are Built Without Frameworks (Primitives over Frameworks)",
    "url": "https://ai.engineer/talks/why-the-best-ai-agents-are-built-without-frameworks-primitives-over-frameworks",
    "videoId": "fcPUqxfrE6Y",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1626000,
    "speakers": [
      {
        "slug": "ahmad-awais",
        "name": "Ahmad Awais",
        "organization": "Langbase"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Ahmad Awais argues that production AI agents should be composed from focused infrastructure primitives instead of heavyweight frameworks. Using Chai and Langbase, he demonstrates a PDF-question-answering agent with memory and discusses composable agent infrastructure, cooperating JavaScript/TypeScript agents, and LLM-as-a-judge evaluation."
  },
  {
    "slug": "why-tts-models-now-look-like-llms-samuel-humeau-mistral",
    "title": "Why TTS Models Now Look Like LLMs — Samuel Humeau, Mistral",
    "url": "https://ai.engineer/talks/why-tts-models-now-look-like-llms-samuel-humeau-mistral",
    "videoId": "3jGAU2sbAyY",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1346000,
    "speakers": [
      {
        "slug": "samuel-humeau",
        "name": "Samuel Humeau",
        "organization": "Mistral AI"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      }
    ],
    "summary": "Mistral AI scientist Samuel Humeau explains contemporary text-to-speech systems using Voxtral TTS, demonstrating voice cloning, multilingual synthesis, and a voice-agent pipeline combining speech recognition, an LLM, and streaming speech generation. He contrasts audio bitrate with the information rate of spoken text to motivate efficient speech representations, then discusses perceived latency, streaming text input, and audience questions about simultaneous text/audio generation and open voice-cloning components."
  },
  {
    "slug": "why-we-don-t-need-more-data-centers",
    "title": "Why We Don’t Need More Data Centers",
    "url": "https://ai.engineer/talks/why-we-don-t-need-more-data-centers",
    "videoId": "M6Vbaig1TsM",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 858000,
    "speakers": [
      {
        "slug": "dr-jasper-zhang",
        "name": "Dr. Jasper Zhang",
        "organization": "Hyperbolic"
      }
    ],
    "topics": [
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Hyperbolic co-founder and CEO Jasper Zhang argues that building additional data centers alone cannot meet growing AI compute demand because construction is costly, energy-intensive, and constrained while existing GPUs remain underused across fragmented providers. He proposes a GPU marketplace and distributed orchestration layer that aggregates available capacity, supports flexible rentals, and lowers acquisition costs. He describes Hyperbolic’s Hyper-dOS software, Kubernetes and MicroK8s integration, H100 pricing examples, and a Monarch-server architecture that coordinates provider provisioning and SSH access. The recording concludes with a question from an unidentified additional participant."
  },
  {
    "slug": "why-we-killed-our-multi-agent-pipeline-subbiah-sethuraman-and-abhilash-asokan-zs-associates",
    "title": "Why We Killed Our Multi-Agent Pipeline — Subbiah Sethuraman and Abhilash Asokan, ZS Associates",
    "url": "https://ai.engineer/talks/why-we-killed-our-multi-agent-pipeline-subbiah-sethuraman-and-abhilash-asokan-zs-associates",
    "videoId": "u6jJcIFDLE4",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 900000,
    "speakers": [
      {
        "slug": "subbiah-sethuraman",
        "name": "Subbiah Sethuraman",
        "organization": "ZS"
      },
      {
        "slug": "abhilash-asokan",
        "name": "Abhilash Asokan",
        "organization": "ZS"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      }
    ],
    "summary": "ZS presenters Subbiah Sethuraman and Abhilash Asokan explain why a pharmaceutical commercial-analytics workflow fragmented across multiple agents produced weak decisions: responsibility for reasoning and shared business context was divided. Their revised architecture detects and prioritizes prescription-related signals deterministically before agent execution, assigns end-to-end reasoning to one agent, and uses a knowledge graph linking geographic entities, payers, brands, and KPIs as a control plane for localization and causal investigation."
  },
  {
    "slug": "why-you-should-care-about-ai-interpretability",
    "title": "Why you should care about AI interpretability",
    "url": "https://ai.engineer/talks/why-you-should-care-about-ai-interpretability",
    "videoId": "6AVMHZPjpTQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1271000,
    "speakers": [
      {
        "slug": "mark-bissell",
        "name": "Mark Bissell",
        "organization": "Goodfire"
      }
    ],
    "topics": [
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "vision-and-video",
        "name": "Vision and video"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Goodfire researcher Mark Bissell explains mechanistic interpretability as reverse-engineering neural networks to inspect and steer meaningful internal features. Using Anthropic's Golden Gate Claude, model privacy controls, Goodfire's Ember and Rakuten's multilingual PII detection, he contrasts direct model interventions with costly LLM-as-a-judge pipelines and fine-tuning. He also discusses interpretability-driven image interfaces, potential applications in biology and genomics, and sparse autoencoders as one approach to discovering interpretable features."
  },
  {
    "slug": "why-your-agent-disagrees-with-itself-and-what-to-do-about-it",
    "title": "Why Your Agent Disagrees With Itself (And What To Do About It)",
    "url": "https://ai.engineer/talks/why-your-agent-disagrees-with-itself-and-what-to-do-about-it",
    "videoId": "wEc9aG7cRQc",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1538000,
    "speakers": [
      {
        "slug": "diane-lin",
        "name": "Diane Lin",
        "organization": "Datadog"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Datadog's Diane Lin explains why AI agents can return contradictory decisions on identical inputs, especially when cybersecurity alerts fall near ambiguous decision boundaries. She argues that disagreement often exposes unclear labels, missing context, or conflicting policies rather than a broken model, and describes prioritizing human review while using semantic knowledge and similar past decisions to make agents more consistent."
  },
  {
    "slug": "why-your-agent-s-brain-needs-a-playbook-practical-wins-from-using-ontologies",
    "title": "Why Your Agent’s Brain Needs a Playbook: Practical Wins from Using Ontologies",
    "url": "https://ai.engineer/talks/why-your-agent-s-brain-needs-a-playbook-practical-wins-from-using-ontologies",
    "videoId": "CbiR9xS2skQ",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 834000,
    "speakers": [
      {
        "slug": "jesus-barrasa",
        "name": "Jesús Barrasa",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "finance",
        "name": "Finance"
      }
    ],
    "summary": "Neo4j AI Field CTO Jesús Barrasa explains how ontologies make GraphRAG applications more structured and reliable by defining domain entities and relationships, guiding knowledge-graph construction from structured or unstructured data, and enriching retrieval. He demonstrates property graphs, schema-driven extraction, vector search, the FIBO ontology, and Cypher-based contextualizing relationships using Neo4j’s GraphRAG Python tooling."
  },
  {
    "slug": "why-your-agents-need-decision-traces-not-just-documents-zach-blumenfeld-neo4j",
    "title": "Why your agents need decision traces, not just documents — Zach Blumenfeld, Neo4j",
    "url": "https://ai.engineer/talks/why-your-agents-need-decision-traces-not-just-documents-zach-blumenfeld-neo4j",
    "videoId": "B9h9ovW5H9U",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1212034,
    "speakers": [
      {
        "slug": "zach-blumenfeld",
        "name": "Zach Blumenfeld",
        "organization": "Neo4j"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      }
    ],
    "summary": "Neo4j’s Zach Blumenfeld explains how context graphs extend document-oriented RAG with connected decision traces and short-term, long-term, and reasoning memory, helping agents act on prior decisions rather than merely retrieve facts. He demonstrates a Neo4j-backed stack using Claude, OpenAI embeddings, Next.js, Graph Data Science, and create-context-graph, then covers Pydantic AI and other agent frameworks, SaaS connectors, MCP, customizable ontologies, and Neo4j Agent Memory. Audience questions explore causal chains and whether decision traces can be assessed or scored for quality."
  },
  {
    "slug": "why-your-ai-ux-is-broken-and-it-s-not-the-model-s-fault",
    "title": "Why Your AI UX Is Broken (and It's Not the Model's Fault)",
    "url": "https://ai.engineer/talks/why-your-ai-ux-is-broken-and-it-s-not-the-model-s-fault",
    "videoId": "YNJvm7t3yq8",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1118000,
    "speakers": [
      {
        "slug": "mike-christensen",
        "name": "Mike Christensen",
        "organization": "Ably"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "Ably staff engineer Mike Christensen explains why connection-bound HTTP and server-sent event streams produce fragile AI chat experiences: disconnects interrupt responses, other tabs lack visibility, and resumability conflicts with cancellation. He presents durable, publish/subscribe-backed sessions and Ably AI Transport as an alternative supporting reconnection, shared multi-device state, agent interaction, and human support handoffs with preserved conversation history."
  },
  {
    "slug": "why-your-enterprise-tech-stack-isn-t-ready-for-ai-agents-and-what-to-build-instead",
    "title": "Why Your Enterprise Tech Stack Isn't Ready for AI Agents - And What to Build Instead",
    "url": "https://ai.engineer/talks/why-your-enterprise-tech-stack-isn-t-ready-for-ai-agents-and-what-to-build-instead",
    "videoId": "mav15aW9lLM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1155000,
    "speakers": [
      {
        "slug": "christopher-lovejoy",
        "name": "Christopher Lovejoy",
        "organization": "Anthropic"
      },
      {
        "slug": "saul-howard",
        "name": "Saul Howard",
        "organization": "Anterior"
      }
    ],
    "topics": [
      {
        "slug": "healthcare",
        "name": "Healthcare"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Christopher Lovejoy of Anthropic and Saul Howard of Anterior explain why successful healthcare AI-agent proofs of concept encounter production barriers around enterprise integrations, compliance-grade audit trails, protected health information, access controls, customer-controlled infrastructure, and clinician escalation. They describe the application, control, and data planes involved in deployment and advocate infrastructure that supports both human and model agents with consistent context and oversight."
  },
  {
    "slug": "why-your-product-needs-an-ai-product-manager-and-why-it-should-be-you",
    "title": "Why your product needs an AI product manager, and why it should be you",
    "url": "https://ai.engineer/talks/why-your-product-needs-an-ai-product-manager-and-why-it-should-be-you",
    "videoId": "xzJdSi2Tsqw",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1117000,
    "speakers": [
      {
        "slug": "james-lowe",
        "name": "James Lowe",
        "organization": "Incubator for Artificial Intelligence (i.AI)"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "industry-applications",
        "name": "Industry applications"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "James Lowe, then Head of AI Engineering at the UK government's Incubator for AI, argues that AI products require product leadership grounded in technical expertise, user needs, experimentation, evaluation, and probabilistic-system safeguards. Drawing on government consultation analysis, rapid prototyping, feature pruning, and ministerial information workflows, he encourages engineers to adopt an AI product-management mindset and validate products with real users."
  },
  {
    "slug": "will-agent-evaluation-via-mcp-stabilize-agent-networks",
    "title": "Will Agent Evaluation via MCP Stabilize Agent Networks?",
    "url": "https://ai.engineer/talks/will-agent-evaluation-via-mcp-stabilize-agent-networks",
    "videoId": "RVN9HWKmkNU",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 851000,
    "speakers": [
      {
        "slug": "ari-heljakka",
        "name": "Ari Heljakka",
        "organization": "Root Signals"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Root Signals CEO Ari Heljakka demonstrates how Model Context Protocol can connect agents to persistent evaluators and judges, creating feedback loops that measure output quality and improve behavior. Examples include using Cursor to optimize marketing copy and adding an MCP-accessible policy evaluator to a Pydantic AI hotel-reservation agent, preventing it from recommending a competing hotel."
  },
  {
    "slug": "windsurf-everywhere-doing-everything-all-at-once",
    "title": "Windsurf everywhere, doing everything, all at once",
    "url": "https://ai.engineer/talks/windsurf-everywhere-doing-everything-all-at-once",
    "videoId": "JVuNPL5QO8Q",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 963000,
    "speakers": [
      {
        "slug": "kevin-hou",
        "name": "Kevin Hou",
        "organization": "Windsurf"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Windsurf head of product Kevin Hou describes an agentic software-development environment built around a shared human-AI timeline. He explains how context from GitHub and MCP-connected services such as Notion, Linear, and Stripe can support work beyond the IDE, including terminal execution with approval controls, browser inspection, pull-request creation, and continuously running parallel agents. He also discusses evaluating end-to-end engineering tasks against unit tests."
  },
  {
    "slug": "wisdom-driven-knowledge-augmented-generation-at-scale",
    "title": "Wisdom-Driven Knowledge Augmented Generation at Scale",
    "url": "https://ai.engineer/talks/wisdom-driven-knowledge-augmented-generation-at-scale",
    "videoId": "9AQOvT8LnMI",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1123000,
    "speakers": [
      {
        "slug": "chin-keong-lam",
        "name": "Chin Keong Lam",
        "organization": "Patho.ai"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Patho.ai founder Chin Keong Lam presents knowledge-augmented generation as a graph-structured alternative to vector-only RAG for expert advisory systems. He explains how knowledge, experience, insight, and feedback contribute to a wisdom-oriented decision framework, demonstrates a competitive-analysis chatbot, and describes prototyping agent orchestration with n8n across OpenAI, Anthropic, and on-premises models. He also discusses quantitative reasoning and points to Neo4j graph-building tools for converting unstructured text into knowledge graphs."
  },
  {
    "slug": "writing-principles-for-task-tuned-prompt-engineering",
    "title": "Writing Principles for Task-Tuned Prompt Engineering",
    "url": "https://ai.engineer/talks/writing-principles-for-task-tuned-prompt-engineering",
    "videoId": "6d60zVdcCV4",
    "event": "AI Engineer Summit 2023",
    "durationMs": 3331000,
    "speakers": [
      {
        "slug": "karina-nguyen",
        "name": "Karina Nguyen",
        "organization": "Anthropic"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "Anthropic engineer Karina Nguyen presents a practical workshop on writing task-tuned prompts for Claude. She explains prompting through next-token prediction, attention, clear instructions, and iterative experimentation, compares Claude with GPT-based systems, and discusses hallucination reduction, faithfulness, scratchpads, formatting, clustering and classification, evaluation, and model self-revision through extensive audience questions."
  },
  {
    "slug": "wtf-do-people-use-open-models-for",
    "title": "WTF do people use Open Models for??",
    "url": "https://ai.engineer/talks/wtf-do-people-use-open-models-for",
    "videoId": "wJwTlvb_TSo",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1681000,
    "speakers": [
      {
        "slug": "eugene-cheah",
        "name": "Eugene Cheah",
        "organization": "Featherless.ai"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "Featherless.ai CEO Eugene Cheah examines how people and businesses actually use open-weight models, contrasting individual demand for DeepSeek-R1, Llama, and Qwen with enterprise reliance on stable, permissively licensed Mistral NeMo deployments. He surveys companionship, creative writing, coding, retrieval, and agent applications, emphasizing model stability, token consumption, production reliability, and future memory-oriented linear-transformer architectures. A second unidentified speaking cluster appears near the closing segment."
  },
  {
    "slug": "wtf-is-the-context-layer-the-missing-infrastructure-for-production-agents",
    "title": "WTF Is the Context Layer? The Missing Infrastructure for Production Agents",
    "url": "https://ai.engineer/talks/wtf-is-the-context-layer-the-missing-infrastructure-for-production-agents",
    "videoId": "8G_1-3IO4ZQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1254000,
    "speakers": [
      {
        "slug": "prukalpa-sankar",
        "name": "Prukalpa Sankar",
        "organization": "Atlan"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Atlan co-founder Prukalpa Sankar argues that production AI agents fail when they lack the business definitions, procedural expertise, and organizational norms human colleagues acquire through experience. Using a fictional analyst, Atlan’s jobs-to-be-done approach, and disconnected marketing and sales agents, she explains why shared contextual infrastructure is needed to coordinate agents and address stale skills, ownership, security, and governance."
  },
  {
    "slug": "you-can-t-prompt-the-room-the-last-skill-ai-won-t-replace",
    "title": "You Can't Prompt the Room: The Last Skill AI Won't Replace",
    "url": "https://ai.engineer/talks/you-can-t-prompt-the-room-the-last-skill-ai-won-t-replace",
    "videoId": "6bmM45jkMDY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 945000,
    "speakers": [
      {
        "slug": "balazs-horvath",
        "name": "Balázs Horváth",
        "organization": "VisualLabs"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      }
    ],
    "summary": "VisualLabs founder Balázs Horváth argues that AI makes identifying valuable problems, engaging stakeholders, and eliciting requirements more important than writing code. He describes a hackathon that rejected 17 of 21 agent ideas over inadequate business value or data access, then presents story mapping, structured user stories, a four-question value framework, and the Value → Architecture → Design sequence as practical tools for moving AI agents beyond demos into genuinely adopted production systems."
  },
  {
    "slug": "you-didn-t-ship-a-bug-you-just-wrote-it-for-a-human",
    "title": "You Didn't Ship a Bug. You Just Wrote It for a Human.",
    "url": "https://ai.engineer/talks/you-didn-t-ship-a-bug-you-just-wrote-it-for-a-human",
    "videoId": "lMCxVorb9wM",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 770000,
    "speakers": [
      {
        "slug": "ravi-madabhushi",
        "name": "Ravi Madabhushi",
        "organization": "Scalekit"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Scalekit cofounder Ravi Madabhushi explains how a human-oriented last-seen timestamp caused recurring latency spikes and database-write pressure when AI agents accessed APIs much faster than people. He contrasts traditional API keys, service accounts, OAuth, and SPIFFE with the distinct identity and authorization requirements of non-deterministic agents, arguing for explicit principal-versus-actor separation, delegated user authorization, fine-grained and time-limited permissions, user-scoped MCP tool access, and visibility into agent actions."
  },
  {
    "slug": "you-might-not-need-50-diffusion-steps",
    "title": "You Might Not Need 50 Diffusion Steps",
    "url": "https://ai.engineer/talks/you-might-not-need-50-diffusion-steps",
    "videoId": "gHs5ZiY80PM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1126000,
    "speakers": [
      {
        "slug": "ziv-ilan",
        "name": "Ziv Ilan",
        "organization": "NVIDIA"
      }
    ],
    "topics": [
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "NVIDIA’s Ziv Ilan explains how diffusion-model inference can move beyond the usual 20–50 denoising steps toward real-time image and video generation. He outlines a stack of quantization, caching, and distillation; discusses post-training versus quantization-aware training, Black Forest Labs’ FLUX.2 on Blackwell GPUs, TensorRT-LLM visual-generation tooling, and chunk-based caching that avoids recomputing unchanged regions. An audience question addresses GB200 access, fine-tuning requirements, and dataset size."
  },
  {
    "slug": "your-agent-architecture-has-a-half-life-of-6-months",
    "title": "Your agent architecture has a half-life of 6 months",
    "url": "https://ai.engineer/talks/your-agent-architecture-has-a-half-life-of-6-months",
    "videoId": "X1kp-ABIIxQ",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1160000,
    "speakers": [
      {
        "slug": "dan-farrelly",
        "name": "Dan Farrelly",
        "organization": "Inngest"
      }
    ],
    "topics": [
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Inngest CTO and co-founder Dan Farrelly argues that rapidly changing models, frameworks and agent patterns demand architectures built around stable execution primitives rather than transient abstractions. He describes durable, resumable execution; human-in-the-loop and sub-agent orchestration; sandboxing; end-to-end observability; and operational triage, emphasizing agent evaluation through measurable engineering and research outcomes."
  },
  {
    "slug": "your-agent-didn-t-fail-your-harness-did",
    "title": "Your Agent Didn’t Fail. Your Harness Did.",
    "url": "https://ai.engineer/talks/your-agent-didn-t-fail-your-harness-did",
    "videoId": "BInpv7lGp1o",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1106000,
    "speakers": [
      {
        "slug": "vinoth-govindarajan",
        "name": "Vinoth Govindarajan",
        "organization": "OpenAI"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Vinoth Govindarajan of OpenAI argues that many production-agent incidents originate in the surrounding harness rather than the model. Using OpenClaw examples, he examines missing durable state, concurrent-writer races, indefinitely stalled runs, improperly scoped approvals, and tool successes that fail to reach users. His recommended architecture emphasizes single-owner state, ordered mutations, deadlines and cancellation, scoped authority, and auditable receipts with external execution evidence and idempotency keys; he closes by recommending the OpenAI Agents SDK."
  },
  {
    "slug": "your-agent-evolved-your-evals-didn-t",
    "title": "Your Agent Evolved. Your Evals Didn't.",
    "url": "https://ai.engineer/talks/your-agent-evolved-your-evals-didn-t",
    "videoId": "nxokqOq1imY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1453000,
    "speakers": [
      {
        "slug": "ameya-bhatawdekar",
        "name": "Ameya Bhatawdekar",
        "organization": "Braintrust"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      }
    ],
    "summary": "Braintrust Field CTO Ameya Bhatawdekar explains why AI-agent evaluations must change as systems evolve from single-prompt applications and RAG chains through ReAct loops, orchestrated workflow graphs, and more capable agent loops. Using an SRE incident-response agent with read and write tools, he highlights new failure surfaces such as node contracts, orchestration, and tool calling, argues for evaluating distributions of outcomes rather than isolated runs, and describes a production-to-evaluation improvement flywheel in Braintrust."
  },
  {
    "slug": "your-agent-failed-in-prod-good-luck-reproducing-it",
    "title": "Your Agent Failed in Prod. Good Luck Reproducing It.",
    "url": "https://ai.engineer/talks/your-agent-failed-in-prod-good-luck-reproducing-it",
    "videoId": "Lc8zRh9muoY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 850000,
    "speakers": [
      {
        "slug": "tisha-chawla",
        "name": "Tisha Chawla",
        "organization": "Microsoft"
      },
      {
        "slug": "susheem-koul",
        "name": "Susheem Koul",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Microsoft engineers Tisha Chawla and Susheem Koul explain why production-agent failures disappear when prompts are rerun, including nondeterminism associated with model architectures, GPU execution, and request batching. They advocate recording semantic inputs and outputs at each agent boundary, then demonstrate Chronicle replaying an erroneous stock-order workflow, inspecting model and tool-call metadata, and converting captured traces into deterministic offline tests and assertions without additional model calls."
  },
  {
    "slug": "your-agent-is-an-infinite-canvas",
    "title": "Your Agent Is an Infinite Canvas",
    "url": "https://ai.engineer/talks/your-agent-is-an-infinite-canvas",
    "videoId": "LMbeDEQO6QM",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1388000,
    "speakers": [
      {
        "slug": "rachel-lee-nabors-rl-nabors",
        "name": "Rachel Lee Nabors (RL Nabors)",
        "organization": "Arize"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Rachel Lee Nabors presents the browser as an interactive canvas for AI agents, using a personal web-comic archive and an MCP social application to demonstrate richer alternatives to chat-only interfaces. The talk contrasts hosted MCP tools and transports, interactive MCP Apps, and WebMCP browser tools; covers edge deployment, CSP/CORS restrictions, navigation callbacks, and existing web APIs; and explains how agent-accessible websites can expose structured actions without relying solely on interface interpretation."
  },
  {
    "slug": "your-agent-is-blindfolded",
    "title": "Your agent is blindfolded",
    "url": "https://ai.engineer/talks/your-agent-is-blindfolded",
    "videoId": "iRcX54EO5g8",
    "event": "AI Engineer Europe 2026",
    "durationMs": 598000,
    "speakers": [
      {
        "slug": "johan-lajili",
        "name": "Johan Lajili",
        "organization": "Poolside"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Poolside engineer Johan Lajili argues that coding-agent reliability depends less on whether software is greenfield or brownfield than on whether agents can inspect applications and verify their own work. He describes Spoolside, an internal CLI that gives agents screenshots, compressed application snapshots, and interaction capabilities for a VS Code extension, creating feedback loops that improve developer trust. He recommends building product-specific agent interfaces as CLIs, skills, or MCP integrations and describes an emerging AIX engineering role focused on making AI-generated work observable and verifiable."
  },
  {
    "slug": "your-agent-is-wasting-tokens-and-you-don-t-know-it",
    "title": "Your Agent Is Wasting Tokens and You Don't Know It",
    "url": "https://ai.engineer/talks/your-agent-is-wasting-tokens-and-you-don-t-know-it",
    "videoId": "uiP88SpCi1Q",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 355000,
    "speakers": [
      {
        "slug": "redacted-username",
        "name": "[REDACTED:username]",
        "organization": "Amazon Web Services"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      }
    ],
    "summary": "AWS developer advocate [REDACTED:username] presents five practical ways to reduce AI-agent token costs using Strands Agents: cache prompts, route requests between cheaper and more capable models, offload and summarize large tool results, cap and observe repeated tool calls, and limit multi-turn conversation history with a sliding-window conversation manager."
  },
  {
    "slug": "your-agent-s-biggest-lie-i-searched-the-web",
    "title": "Your Agent's Biggest Lie: \"I Searched the Web\"",
    "url": "https://ai.engineer/talks/your-agent-s-biggest-lie-i-searched-the-web",
    "videoId": "btxGmN8RvNU",
    "event": "AI Engineer Europe 2026",
    "durationMs": 949000,
    "speakers": [
      {
        "slug": "rafael-levi",
        "name": "Rafael Levi",
        "organization": "Bright Data"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Bright Data’s Rafael Levi explains how AI agents can present blocked, empty, or stale web-access attempts as successful searches, creating fabricated citations and unreliable answers. He demonstrates Bright Data’s Web MCP against public pages on LinkedIn, Instagram, Amazon, TikTok, and property websites; discusses CAPTCHAs, Cloudflare AI Labyrinth, public-data access boundaries, existing datasets, agent skills, generated scraping pipelines, and MCP tool-selection overhead; and answers audience questions."
  },
  {
    "slug": "your-agents-need-a-save-button",
    "title": "Your Agents Need a Save Button",
    "url": "https://ai.engineer/talks/your-agents-need-a-save-button",
    "videoId": "bZISsg7H7DA",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1027000,
    "speakers": [
      {
        "slug": "hamza-tahir",
        "name": "Hamza Tahir",
        "organization": "ZenML"
      }
    ],
    "topics": [
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "ZenML co-founder Hamza Tahir argues that conventional agent traces fail to preserve executable runtime state and demonstrates Kitaru as a durable checkpointing and replay layer. He shows how to snapshot artifacts and environments, replay executions with a cheaper model, compare alternative runs, and use JSON reports and an MCP server to analyze agent behavior across cohorts rather than relying on isolated examples."
  },
  {
    "slug": "your-ai-agent-isn-t-an-engineer-the-art-of-thoughtful-anthropomorphism",
    "title": "Your AI Agent Isn't an Engineer: The Art of Thoughtful Anthropomorphism",
    "url": "https://ai.engineer/talks/your-ai-agent-isn-t-an-engineer-the-art-of-thoughtful-anthropomorphism",
    "videoId": "MExbNNG_VcI",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1220000,
    "speakers": [
      {
        "slug": "rizel-scarlett",
        "name": "Rizel Scarlett",
        "organization": "Block"
      }
    ],
    "topics": [
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Rizel Scarlett, then a staff developer advocate at Block, argues that marketing AI agents as human software engineers undermines developer trust, encourages replacement narratives, and creates unrealistic expectations. Using examples including Claude, Devin, ChatGPT, and the 2024 Stack Overflow Developer Survey, she advocates thoughtful anthropomorphism, authentic developer marketing, hands-on customer-zero product experience, and positioning AI as a tool that supports rather than replaces developers."
  },
  {
    "slug": "your-ai-product-will-fail-unless-you-can-explain-it",
    "title": "Your AI Product Will Fail Unless You Can Explain It",
    "url": "https://ai.engineer/talks/your-ai-product-will-fail-unless-you-can-explain-it",
    "videoId": "d_Ftrl3vfV0",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 362000,
    "speakers": [
      {
        "slug": "veronica-hylak",
        "name": "Veronica Hylak",
        "organization": "Hey AI"
      }
    ],
    "topics": [
      {
        "slug": "enterprise",
        "name": "Enterprise"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      }
    ],
    "summary": "Veronica Hylak of Hey AI explains how founders can replace opaque agentic-AI pitches with memorable, customer-centered product stories. She recommends starting with a concrete customer pain point, making the product understandable through familiar examples and visual analogies, and describing specific changes in the user’s workflow. Examples include fragmented security operations, AI-agent observability, Devin, and information scattered across Slack and spreadsheets."
  },
  {
    "slug": "your-attention-is-the-bottleneck-not-your-agents-zack-proser-workos",
    "title": "Your Attention Is the Bottleneck, Not Your Agents — Zack Proser, WorkOS",
    "url": "https://ai.engineer/talks/your-attention-is-the-bottleneck-not-your-agents-zack-proser-workos",
    "videoId": "so9l_MwS2yg",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1517000,
    "speakers": [
      {
        "slug": "zack-proser",
        "name": "Zack Proser",
        "organization": "WorkOS"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "WorkOS Applied AI engineer Zack Proser argues that human attention, rather than agent throughput, limits sustainable AI-assisted software development. Using an internal BlogBot bug as an example, he shows how Claude Code, Slack, Linear, and MCP can complete and verify a fix without repeated human intervention. He recommends voice-first parallel workflows, layered lint/build/test and browser-verification gates, continuous improvement from agent conversation histories, and attention to developer well-being, including an Oura Ring MCP integration. Audience questions address early-career skill development, unwieldy JSONL histories, and voice interaction."
  },
  {
    "slug": "your-coding-agent-doesn-t-always-follow-your-rules",
    "title": "Your coding agent doesn't always follow your rules",
    "url": "https://ai.engineer/talks/your-coding-agent-doesn-t-always-follow-your-rules",
    "videoId": "MpZzWMdmQCE",
    "event": "AI Engineer Europe 2026",
    "durationMs": 608000,
    "speakers": [
      {
        "slug": "talha-sheikh",
        "name": "Talha Sheikh",
        "organization": "Checkout.com"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "evals",
        "name": "Evals"
      }
    ],
    "summary": "Talha Sheikh of Checkout.com explains why coding agents can claim completion while missing requirements, leaving humans to enforce correctness. He introduces Vector Harness, which uses Claude Code hooks and deterministic checks to verify agent output, and discusses multi-agent workflows, asynchronous checks, LLM-as-a-judge evaluations, code-review feedback loops, and the broader shift toward harness engineering."
  },
  {
    "slug": "your-coding-agent-just-got-cloned-and-your-brain-isn-t-ready",
    "title": "Your Coding Agent Just Got Cloned And Your Brain Isn't Ready",
    "url": "https://ai.engineer/talks/your-coding-agent-just-got-cloned-and-your-brain-isn-t-ready",
    "videoId": "X4BwOu0GWb8",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 820000,
    "speakers": [
      {
        "slug": "rustin-banks",
        "name": "Rustin Banks",
        "organization": "Google Labs"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Google Labs product manager Rustin Banks presents Jules, an asynchronous coding agent that runs tasks concurrently against GitHub repositories in cloud virtual machines. He contrasts serial programming with orchestrating parallel agents, discusses task generation and downstream review and merging, and demonstrates competing Jest and Playwright implementations alongside conference-app features, accessibility audits, and security audits. He concludes that richer context improves agent performance and identifies Gemini 2.5 Pro as Jules' underlying model at the time of the talk."
  },
  {
    "slug": "your-coding-agent-should-do-ai-system-engineering",
    "title": "Your Coding Agent Should Do AI System Engineering",
    "url": "https://ai.engineer/talks/your-coding-agent-should-do-ai-system-engineering",
    "videoId": "JomVvNDjGb8",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1105000,
    "speakers": [
      {
        "slug": "ben-burtenshaw",
        "name": "Ben Burtenshaw",
        "organization": "Hugging Face"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Hugging Face’s Ben Burtenshaw argues that coding agents can extend beyond application programming into AI systems engineering through three increasingly autonomous workflows: writing and distributing CUDA kernels, fine-tuning language models from a prompt, and coordinating multi-agent research experiments. He explains GPU memory bottlenecks and FlashAttention, introduces Hugging Face’s Kernels repository, compares skill-enabled model performance, and describes Git-based experiment management with dashboard and Parquet access."
  },
  {
    "slug": "your-evals-are-meaningless-and-here-s-how-to-fix-them",
    "title": "Your Evals Are Meaningless (And Here’s How to Fix Them)",
    "url": "https://ai.engineer/talks/your-evals-are-meaningless-and-here-s-how-to-fix-them",
    "videoId": "3jwClx0Ft2E",
    "event": "AI Engineer Summit 2025",
    "durationMs": 1130000,
    "speakers": [
      {
        "slug": "mohak-sharma",
        "name": "Mohak Sharma",
        "organization": "HoneyHive"
      }
    ],
    "topics": [
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "HoneyHive co-founder Mohak Sharma explains why developer-written test sets, generic metrics, and uncalibrated LLM judges can produce evaluation scores that fail to predict real production performance. He recommends domain-expert-defined criteria and edge cases, continuously incorporating problematic production queries into evaluation datasets, and tracking agreement between automated evaluators and human judgments using F1 scores or correlation metrics."
  },
  {
    "slug": "your-finance-agent-s-bottleneck-is-you",
    "title": "Your Finance Agent's Bottleneck Is You",
    "url": "https://ai.engineer/talks/your-finance-agent-s-bottleneck-is-you",
    "videoId": "z0sh8HyTrDo",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 822000,
    "speakers": [
      {
        "slug": "ramana-siddanth-emani",
        "name": "Ramana Siddanth Emani",
        "organization": "Auditoria AI"
      }
    ],
    "topics": [
      {
        "slug": "finance",
        "name": "Finance"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      }
    ],
    "summary": "Auditoria AI data scientist Siddanth Emani argues that finance-agent deployments are constrained less by models or hardware than by developers' ability to diagnose production failures and iterate quickly. He proposes parallel subagents isolated in Git worktrees, MCP-connected tools, and an automated workflow spanning Jira tickets, root-cause analysis, test-driven fixes, pull requests, Docker deployments, and QA. A unified interface brings together Kubernetes, Jira, GitHub, and Claude Code while preserving finance controls and positioning humans as verifiers rather than throughput bottlenecks."
  },
  {
    "slug": "your-fine-tuned-model-is-tech-debt-a-50x-roi-house-of-cards",
    "title": "Your Fine-Tuned Model Is Tech Debt: A 50x ROI House of Cards",
    "url": "https://ai.engineer/talks/your-fine-tuned-model-is-tech-debt-a-50x-roi-house-of-cards",
    "videoId": "4loPnxvWWhg",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 999020,
    "speakers": [
      {
        "slug": "dan-bjornn",
        "name": "Dan Bjornn",
        "organization": "Lease End"
      }
    ],
    "topics": [
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      }
    ],
    "summary": "Lease End senior data scientist Dan Bjornn describes replacing a retrieval-based customer-intent workflow with a fine-tuned model that reportedly generated $12 million in revenue at 50x ROI, while accumulating operational debt through labeling, synthetic-example generation, retraining, customer-facing classification failures, and vendor lock-in. His team rebuilt around a model-agnostic agentic framework, deployed changes as Markdown files in Amazon S3, accepted higher per-message API costs, and regained flexibility to use providers including OpenAI and Anthropic. He argues that fine-tuning should be a last resort when frontier models cannot solve the problem."
  },
  {
    "slug": "your-insecure-mcp-server-won-t-survive-production-tun-shwe-lenses",
    "title": "Your Insecure MCP Server Won't Survive Production — Tun Shwe, Lenses",
    "url": "https://ai.engineer/talks/your-insecure-mcp-server-won-t-survive-production-tun-shwe-lenses",
    "videoId": "BurJvbqFr4c",
    "event": "AI Engineer Europe 2026",
    "durationMs": 1474000,
    "speakers": [
      {
        "slug": "tun-shwe",
        "name": "Tun Shwe",
        "organization": "Lenses.io"
      },
      {
        "slug": "jeremy-frenay",
        "name": "Jeremy Frenay",
        "organization": "Lenses.io"
      }
    ],
    "topics": [
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      }
    ],
    "summary": "Lenses.io engineers Tun Shwe and Jeremy Frenay explain why enterprise MCP security begins with agent-oriented interface design: reduce exposed tools, constrain inputs with validation, defend tool descriptions against poisoning, and avoid leaking sensitive context. Grounding these risks in OWASP guidance and Lenses’ Kafka-focused open-source MCP server, they contrast local development with production requirements and discuss OAuth authentication, weaknesses of long-lived credentials, and scoped token validation."
  },
  {
    "slug": "your-llm-deception-monitor-is-broken-the-fix-is-in-the-training-data-sachin-kumar-lexisnexis",
    "title": "Your LLM Deception Monitor Is Broken. The Fix Is in the Training Data - Sachin Kumar, LexisNexis",
    "url": "https://ai.engineer/talks/your-llm-deception-monitor-is-broken-the-fix-is-in-the-training-data-sachin-kumar-lexisnexis",
    "videoId": "IQkVMvXQKLY",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 838000,
    "speakers": [
      {
        "slug": "sachin-kumar",
        "name": "Sachin Kumar",
        "organization": "LexisNexis"
      }
    ],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "observability-and-reliability",
        "name": "Observability and reliability"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Sachin Kumar of LexisNexis presents independent research on detecting sleeper-agent backdoors that behavioral evaluations and safety fine-tuning can miss. His approach subtracts base-model activations from fine-tuned-model activations and trains a difference-based sparse autoencoder on that signal. In a controlled, year-triggered SQL-injection experiment using SmolLM2-360M, he reports a 40-fold stronger signal than joint cross-model features, perfect precision, zero false positives, and robustness across middle layers and fine-tuning regimes. He proposes integrating the resulting delta monitor into fine-tuning pipelines while noting that a single feature detects approximately 25% of triggers and that the experiment covered one backdoor type on one 360-million-parameter model."
  },
  {
    "slug": "your-llm-ran-out-of-knowledge-now-what",
    "title": "Your LLM Ran Out of Knowledge — Now What?",
    "url": "https://ai.engineer/talks/your-llm-ran-out-of-knowledge-now-what",
    "videoId": "ya_9_niq2as",
    "event": "AI Engineer Summit 2025",
    "durationMs": 771000,
    "speakers": [],
    "topics": [
      {
        "slug": "other-unclassified",
        "name": "Other / unclassified"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "The talk demonstrates an open-source reasoning framework for applying LLMs in domains with little usable training data, including corporate negotiations and geopolitics. It combines structured scenario parsing, domain-specific heuristics, explicit rules, and realistic constraints; a WorldSim-assisted geopolitical example uses Anthropic and eliminates scenarios that violate prescribed rules while retaining human oversight."
  },
  {
    "slug": "your-llm-stack-is-a-2008-database-with-better-marketing",
    "title": "Your LLM Stack Is a 2008 Database With Better Marketing",
    "url": "https://ai.engineer/talks/your-llm-stack-is-a-2008-database-with-better-marketing",
    "videoId": "XjI-AR4pt7Y",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1236000,
    "speakers": [
      {
        "slug": "lovina-dmello",
        "name": "Lovina Dmello",
        "organization": "NVIDIA"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "NVIDIA infrastructure engineer Lovina Dmello argues that production ML and LLM breaches usually arise from familiar infrastructure misconfigurations rather than exotic model attacks. Using exposed Ray clusters as an example, she maps threats across model, data, supply-chain, and infrastructure layers; examines the latency and throughput costs of security controls; and recommends authenticated access, least privilege, short-lived credentials, container isolation, network segmentation, and protection of stored data."
  },
  {
    "slug": "your-mcp-server-is-bad-and-you-should-feel-bad",
    "title": "Your MCP Server is Bad and You Should Feel Bad",
    "url": "https://ai.engineer/talks/your-mcp-server-is-bad-and-you-should-feel-bad",
    "videoId": "96G7FLab8xc",
    "event": "AI Engineer Code 2025",
    "durationMs": 3273000,
    "speakers": [
      {
        "slug": "jeremiah-lowin",
        "name": "Jeremiah Lowin",
        "organization": "Prefect"
      }
    ],
    "topics": [
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      }
    ],
    "summary": "Prefect founder and FastMCP creator Jeremiah Lowin explains why MCP servers should be designed as agent-native products rather than exposed REST APIs. He covers focused tool surfaces, agent-centered workflows, orchestration costs, client behavior including Claude Desktop tool caching, pragmatic API bootstrapping, and SEP-1686 asynchronous background tasks."
  },
  {
    "slug": "your-moat-is-your-data-model",
    "title": "Your Moat Is Your Data Model",
    "url": "https://ai.engineer/talks/your-moat-is-your-data-model",
    "videoId": "jt1Pbr_n6oU",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 1230000,
    "speakers": [
      {
        "slug": "mike-phipps",
        "name": "Mike Phipps",
        "organization": "Gates Foundation"
      }
    ],
    "topics": [
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      }
    ],
    "summary": "Mike Phipps explains why an enterprise AI team's durable advantage lies in modeling its organizational processes and tacit knowledge rather than owning rapidly commoditizing models or interfaces. At the Gates Foundation, the Strategic Intelligence Platform uses a cross-system semantic knowledge graph to support agentic workflows, represent funding and management hierarchies, connect siloed operational data, and enforce governance practices including PII masking. The official session description identifies Neo4j, an MCP server, Claude integrations, and retrieval evaluations as additional elements of the architecture."
  },
  {
    "slug": "your-personal-open-source-humanoid-robot-for-8-999-jingxiang-jx-mo-k-scale-labs",
    "title": "Your Personal Open-Source Humanoid Robot for $8,999 — Jingxiang \"JX\" Mo, K-Scale Labs",
    "url": "https://ai.engineer/talks/your-personal-open-source-humanoid-robot-for-8-999-jingxiang-jx-mo-k-scale-labs",
    "videoId": "BS92RdBvI90",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 1166000,
    "speakers": [
      {
        "slug": "jingxiang-jx-mo",
        "name": "Jingxiang \"JX\" Mo",
        "organization": "K-Scale Labs"
      }
    ],
    "topics": [
      {
        "slug": "robotics-and-world-models",
        "name": "Robotics and world models"
      }
    ],
    "summary": "K-Scale Labs founding engineer Jingxiang \"JX\" Mo presents K-Bot, an approximately $9,000 open-source humanoid, alongside the smaller Z-Bot. He describes openly available hardware and software, modular upgrades, VR teleoperation, Python/Rust development tools, reinforcement-learning locomotion and manipulation, simulation with Isaac Sim or MJX, and support for vision-language and vision-language-action models. The recording concludes with audience questions about practical applications, trade-offs, and competing humanoid prices."
  },
  {
    "slug": "your-realtime-ai-is-ngmi-sean-dubois-openai-kwindla-hultman-kramer-daily",
    "title": "Your realtime AI is ngmi — Sean DuBois (OpenAI), Kwindla Hultman Kramer (Daily)",
    "url": "https://ai.engineer/talks/your-realtime-ai-is-ngmi-sean-dubois-openai-kwindla-hultman-kramer-daily",
    "videoId": "E71YtNbCFXY",
    "event": "AI Engineer World's Fair 2025",
    "durationMs": 990000,
    "speakers": [
      {
        "slug": "sean-dubois",
        "name": "Sean DuBois",
        "organization": "OpenAI"
      },
      {
        "slug": "kwindla-hultman-kramer",
        "name": "Kwindla Hultman Kramer",
        "organization": "Daily"
      },
      {
        "slug": "yaxin",
        "name": "Yaxin"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "architecture",
        "name": "Architecture"
      }
    ],
    "summary": "OpenAI's Sean DuBois and Daily's Kwindla Hultman Kramer explain why voice-AI systems should be designed around low-latency networking, emphasizing conversational response times near 500 milliseconds, end-to-end voice-to-voice latency, and WebRTC's advantages over WebSockets for streaming audio. They discuss the OpenAI Realtime API, Pion, and Pipecat; demonstrate Squabbert, a Raspberry Pi-based voice agent using MLX Whisper and Gemma 3; and include a community project introduction."
  },
  {
    "slug": "your-support-team-should-ship-code",
    "title": "Your Support Team Should Ship Code",
    "url": "https://ai.engineer/talks/your-support-team-should-ship-code",
    "videoId": "RmJ4rTLV_x4",
    "event": "AI Engineer Code 2025",
    "durationMs": 966000,
    "speakers": [
      {
        "slug": "lisa-orr",
        "name": "Lisa Orr",
        "organization": "Zapier"
      }
    ],
    "topics": [
      {
        "slug": "developer-workflows-and-testing",
        "name": "Developer workflows and testing"
      },
      {
        "slug": "leadership",
        "name": "Leadership"
      },
      {
        "slug": "apis-mcp-and-protocols",
        "name": "APIs, MCP, and protocols"
      },
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      }
    ],
    "summary": "Zapier product leader Lisa Orr explains how constantly changing third-party APIs create integration failures and support backlogs, and how support staff can help ship fixes under engineering review. After discovery revealed that gathering documentation, logs, and bug context was a major bottleneck, the team built LLM-assisted diagnosis and testing tools; its standalone Autocode playground struggled because it interrupted existing workflows. Adoption improved by embedding Diagnosis into Zapier and Jira workflows, using MCP alongside tools such as Cursor, and orchestrating the Scout project through Zaps so completed merge requests reach support for validation."
  },
  {
    "slug": "your-voice-agent-doesn-t-need-a-frontier-model",
    "title": "Your Voice Agent Doesn't Need a Frontier Model",
    "url": "https://ai.engineer/talks/your-voice-agent-doesn-t-need-a-frontier-model",
    "videoId": "fnLBmfsI_Fg",
    "event": "AI Engineer World's Fair 2026",
    "durationMs": 345000,
    "speakers": [
      {
        "slug": "joel-allou",
        "name": "Joel Allou"
      },
      {
        "slug": "ornella-bahidika",
        "name": "Ornella Bahidika",
        "organization": "Microsoft"
      }
    ],
    "topics": [
      {
        "slug": "speech-and-audio",
        "name": "Speech and audio"
      },
      {
        "slug": "agent-engineering",
        "name": "Agent engineering"
      },
      {
        "slug": "safety-and-governance",
        "name": "Safety and governance"
      }
    ],
    "summary": "Joel Allou and Ornella Bahidika present Ace, an AI voice tutor designed around an approximately 950-millisecond response budget. Instead of relying on a frontier model, deterministic state machines handle lesson progression, student knowledge, and turn-by-turn context while Claude Haiku 4.5 generates speech. A comparison against Claude Opus 4.7 reports roughly 900-millisecond responses for the scaffolded smaller model, with strict structural guardrails needed to control drift."
  },
  {
    "slug": "z-ai-glm-4-6-what-we-learned-from-100-million-open-source-downloads-yuxuan-zhang-z-ai",
    "title": "Z.ai GLM-4.6: What We Learned From 100 Million Open Source Downloads — Yuxuan Zhang, Z.ai",
    "url": "https://ai.engineer/talks/z-ai-glm-4-6-what-we-learned-from-100-million-open-source-downloads-yuxuan-zhang-z-ai",
    "videoId": "m6MF1OR_9kM",
    "event": "AI Engineer Code 2025",
    "durationMs": 1179000,
    "speakers": [
      {
        "slug": "yuxuan-zhang",
        "name": "Yuxuan Zhang",
        "organization": "Z.ai"
      }
    ],
    "topics": [
      {
        "slug": "coding-and-developer-tools",
        "name": "Coding and developer tools"
      },
      {
        "slug": "evals",
        "name": "Evals"
      },
      {
        "slug": "infrastructure-and-deployment",
        "name": "Infrastructure and deployment"
      },
      {
        "slug": "rag-context-and-search",
        "name": "RAG, context, and search"
      },
      {
        "slug": "data-and-model-adaptation",
        "name": "Data and model adaptation"
      },
      {
        "slug": "reasoning-and-models",
        "name": "Reasoning and models"
      }
    ],
    "summary": "Z.ai engineer Yuxuan Zhang surveys the open-source GLM ecosystem, reporting more than 100 million model downloads, before describing GLM-4.6’s coding and preference-benchmark results. He explains CC-Bench’s realistic coding-agent evaluation, approximately 15 trillion pretraining tokens, and the open-source slime reinforcement-learning framework’s SGLang-based hybrid synchronous/asynchronous architecture. The talk also covers temporal tokens for video understanding and practical access to Z.ai models and open-source deployment resources."
  }
]