# Agentic Sites: Building Hyper Personalized Websites — Carlos Sanchez, Adobe

Carlos Sanchez, Adobe

AI Engineer World's Fair 2026

[Talk page](https://ai.engineer/talks/jebp4V0vh30-agentic-sites-building-hyper-personalized-websites) · [Video](https://www.youtube.com/watch?v=jebp4V0vh30)

## Agentic Sites: Building Hyper-Personalized Websites

Carlos Sanchez shows how Adobe’s prototype turns visitor intent into a purpose-built page by changing selected content blocks, grounding them in the existing site, and treating latency as part of model quality.

### Start with what the visitor is trying to accomplish

[0:13](https://www.youtube.com/watch?v=jebp4V0vh30&t=13s)

Carlos Sanchez begins with a different organizing question for a website: what is this visitor trying to do right now? Instead of serving one fixed page to everyone, the proposed system interprets browsing behavior and explicit requests, then adapts the experience in real time. Engagement and conversion are the intended marketing outcomes, although the talk does not present measurements showing that this implementation improves either one.

The important constraint is that the system does not freely generate an entire website. It changes selected content blocks according to a persona or inferred intent while preserving the surrounding site. That smaller action space fits the practical requirements of brands with strict visual and editorial guidelines: the model can choose, arrange, and rewrite approved material without inventing every component from scratch.

To constrain the content itself, the team turns the existing site into a retrieval corpus. Retrieval-augmented generation supplies relevant products, guides, stories, and other site material as context for each response. This connects generated copy to available content, but grounding alone cannot guarantee that every recommendation or claim will be correct; it narrows the source material rather than eliminating model error.

Sanchez describes three forms of adaptation:

- **Persona adaptation:** change selected blocks as the system classifies the visitor.
- **Query-driven assembly:** build a results page around an explicit search or request.
- **Behavior-based recommendation:** prepare a page from what the visitor has browsed and what the system infers they may want.

The third mode is the least direct because it relies on inferred intent rather than a stated request.

Marketers define the strategy in natural language and use analytics to revise it. The intended loop runs from a business goal to a personalized experience, then from observed behavior back into the next strategy change. Sanchez describes the loop at a conceptual level; he does not specify an optimization algorithm, attribution method, or success threshold.

### A page generator has an interaction budget

[4:30](https://www.youtube.com/watch?v=jebp4V0vh30&t=270s)

The frontend composes adaptable blocks and updates them through the AI service. Behind it, the team continuously evaluates prompts across multiple models and providers. Model selection is deliberately site-specific: corpus size, subject matter, commerce type, and the required outputs can all change which configuration is accurate enough and fast enough.

The team uses Promptfoo for comparisons across local and hosted providers, evaluating accuracy alongside speed. Latency is not an infrastructure footnote here. If generation happens after a visitor asks for a page, inference time becomes page-load time, so Sanchez sets a target of roughly one or two seconds.

For one example site and a set of 15 prompts, the selected Cerebras configuration averaged 1.1 seconds to generate a page; the next configuration averaged 4.6 seconds. That large gap drove the provider choice for the demonstration. These results describe one prompt set and site, not a universal ranking of providers or models.

This produces a practical selection rule: choose enough capability to satisfy the task within the interaction budget. The model is mostly generating bounded copy, selecting blocks, and arranging them, so Sanchez argues that the workload does not require the largest available LLM. A configuration can be imperfect yet useful if its quality is acceptable for the site and its response arrives before the experience feels broken. The talk leaves “good enough” to application-specific evaluation rather than defining a universal threshold.

Model selection is a site-specific tradeoff
The chosen configuration must meet both the site’s quality requirements and the page’s interaction budget.
- Site-specific prompts — Prompts reflect the site’s size, domain, commerce type, and desired adaptations.
- Accuracy evaluation — Outputs must be suitable for the site’s task and content.
- Latency evaluation — On-demand page generation targets roughly one or two seconds.
- Per-site configuration — Select a model and provider that are accurate enough and fast enough for this experience.
- Site-specific prompts → Accuracy evaluation (test output quality)
- Site-specific prompts → Latency evaluation (measure generation time)
- Accuracy evaluation → Per-site configuration (must be good enough)
- Latency evaluation → Per-site configuration (must meet interaction budget)

### Move some generation ahead of the click

[8:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=496s)

The runtime inputs combine browsing activity with explicit queries. In the example, those signals can change the hero card, product selection, blog feed, navigation, and calls to action. Personalization therefore affects emphasis, ordering, recommendations, and the next action—not just a line of greeting text.

A “For You” page creates a useful timing option. As the visitor browses, the backend can repeatedly generate or refresh recommendations and prefetch the resulting page. When the visitor opens it, the expensive work may already be complete. This relaxes the immediate latency requirement, but it spends additional inference before the system knows whether the visitor will use the result.

Freshness creates the corresponding cost. Every new signal can make the prepared recommendation stale, so keeping it aligned with current behavior may require more generations and more LLM calls. The design trades waiting at navigation time for background inference cost.

Queries can also place visitors into marketer-defined personas or intent groups, such as buying a product versus gathering information. Marketers decide which groups exist and how they should be treated; AI selects the blocks, suggestions, sequence, and potentially media for the selected group. This keeps the segmentation strategy under business control while delegating page assembly to the model.

Sanchez treats generated images as a possible extension, with a sharper brand risk than text. Image generation is slower, and marketing acceptance depends on quality and brand fit. The demonstrated architecture does not require it: product pages and comparisons can be assembled from existing products, guides, experiences, and site media.

The stack separates observation, generation, retrieval, and delivery. The browser collects signals. A backend calls the LLM and reasons over retrieval results from the site’s vector database. Adobe Experience Manager serves pages and static content at the edge. This lets the static site remain fast and controlled while selected blocks take a dynamic path through retrieval and inference.

That signal path also creates an important production concern that the presentation does not explore: the demo records visited pages and time spent, then makes those observations available to the LLM. A deployed system would need explicit decisions about consent, retention, access, and what behavioral data may be sent to an inference service.

From visitor behavior to an adapted page
Static site content remains at the edge while selected blocks travel through a retrieval-and-generation path.
- Visitor activity — Browsing and queries provide intent signals.
- Persona or intent group — The system classifies the visitor using groups defined by marketers.
- Site retrieval corpus — A vector database retrieves relevant existing site content.
- LLM assembly — The backend selects, orders, and rewrites adaptable blocks.
- Edge-delivered page — Static content is served with personalized heroes, products, feeds, navigation, and calls to action.
- Visitor activity → Persona or intent group (signals inform classification)
- Persona or intent group → Site retrieval corpus (guides relevant content selection)
- Site retrieval corpus → LLM assembly (grounds the response)
- LLM assembly → Edge-delivered page (updates selected blocks)

### A camping request changes the whole product story

[12:36](https://www.youtube.com/watch?v=jebp4V0vh30&t=756s)

Sanchez calls the ambition “Audience of One”: a marketing experience assembled for an individual rather than a broad segment. His demonstration uses a fully generated example site for coffee equipment. He browses stories and products, spending time on individual pages before opening a debugging panel.

The panel shows the mechanism behind the apparent personalization. The visitor is bucketed as “exploring,” while the system records visited pages and time spent on each. Those signals are available to the LLM, and a “For You” page has already been generated from the browsing history.

The explicit-query demo makes the transformation easier to inspect. Sanchez asks for a coffee machine suitable for preparing coffee while camping. The resulting page changes its introductory copy, adds camping-specific tips, and recommends two machines described as appropriate for that use. The page is more than a ranked search result: several blocks tell one coherent product story around the same intent.

A similar run in debug mode takes 1.64 seconds in total, including the round trip to the LLM. Sanchez reports approximately 2,300 tokens per second for that run; another run shows about one second of LLM time and roughly 2,200 tokens per second. These numbers describe different layers of performance: total generation time includes more than model decoding, while token throughput measures only generation rate.

The demo interface also lets the team switch models and adjust temperature and token settings. Automated Promptfoo tests provide repeatable comparisons, while manual interaction reveals whether a configuration behaves well inside the actual page experience. The two methods answer different questions: batch evaluation measures defined cases; interactive inspection exposes coherence, responsiveness, and presentation problems that a score may miss.

### Carry the pattern to another site in under an hour

[16:40](https://www.youtube.com/watch?v=jebp4V0vh30&t=1000s)

Sanchez then introduces Of One Labs, a tool that accepts a URL and generates an agentic demonstration site in less than an hour. The claim concerns creating a demo, not automatically converting any production site with its full brand, security, analytics, accessibility, and editorial requirements intact.

He applied the tool to the AI Engineer site. The generated experience centers on a search box with AI-generated suggestions; a query for European AI conferences is intended to assemble a page focused on relevant events. As in the coffee example, the query supplies the job the page should perform rather than merely a keyword to match.

The strongest result is a side-by-side conference comparison. Once the system infers that a visitor is choosing between two events, it changes the information structure to support that decision. This is the more consequential form of personalization in the talk: the system does not simply rank one item above another; it selects a comparison layout suited to the visitor’s task.

### Intent-driven assembly can outlive the conventional webpage

[18:17](https://www.youtube.com/watch?v=jebp4V0vh30&t=1097s)

The closing example asks what this pattern might become if a personal assistant mediates the interaction. Sanchez imagines a spoken shopping request producing a personalized visual response on a television. The user supplies intent by voice, and the system chooses a screen-sized presentation without requiring a phone or computer.

He is careful not to claim certainty about whether the web itself is dying or what will replace it. The supported result is narrower: the demonstration shows that a system can already assemble selected, site-grounded blocks around a visitor’s inferred or explicit intent. Extending that result into a general assistant-driven interface remains a proposed direction.

Sanchez expects generation to become faster, cheaper, and more widely applicable, but those are forecasts. The engineering lesson available today is concrete: constrain what may change, retrieve from the site’s approved content, evaluate models against each site’s quality and latency needs, and assemble the page around the job the visitor is trying to complete.

## Key takeaways

- Personalize selected blocks rather than freely generating the whole website; use the existing site as the retrieval corpus so adaptations remain connected to approved content.
- Treat latency as part of model quality. In the reported 15-prompt example, the chosen configuration averaged 1.1 seconds versus 4.6 seconds for the next option, but that comparison does not establish universal provider performance.
- Pre-generating a “For You” page moves inference ahead of navigation, reducing visible waiting at the cost of repeated calls and potentially wasted generation.
- Marketers define the intent groups and strategy; the model chooses and orders suitable blocks for the selected group.
- The camping-coffee and conference-comparison demos show the central capability: reorganize several grounded blocks around a visitor’s task, rather than returning a conventional list of search results.
- Behavioral personalization requires more than model engineering. Because the prototype records visited pages and time spent, a production design also needs explicit privacy and data-governance decisions.

## Resources

- [Agentic Sites: Building Hyper Personalized Websites](https://ai.engineer/talks/jebp4V0vh30-youtube-76887db4d455a40ac4c7) — The official AI Engineer page includes the recording, timestamped transcript, chapters, and a concise talk summary.
- [Carlos Sanchez’s Weblog](https://blog.csanchez.org/) — Sanchez’s technical weblog covers his broader work in AI, infrastructure, Kubernetes, and automation.
- [Carlos Sanchez on X](https://x.com/csanchez) — The speaker’s public profile for updates on his technical projects and presentations.

## Related talks

- [Personalization in the Era of LLMs](https://ai.engineer/talks/5YSJEP0HWzM-personalization-in-era-llms) — A complementary look at LLM-based personalization in recommendation systems, including user representations, catalog embeddings, and steerable music and podcast recommendations.
- [Agent Output Is Not UX: Rendering Layer Your LLM Pipeline Is Missing - Bala Ramdoss, Amazon Lens](https://ai.engineer/talks/maTp79FD9gI-agent-output-is-not-ux-rendering-layer) — Extends the interface question by explaining why model output needs a controlled rendering layer, version-aware contracts, and safe fallbacks before it becomes production UX.
- [How to evaluate a model for your use case](https://ai.engineer/talks/pj_hKFhnJCw-evaluate-model-your-use-case) — Develops the talk’s model-selection principle: generic benchmarks cannot replace application-specific evaluations tied to the actual deployment task.

## Slides

### Agentic Sites: Building Hyper Personalized Websites

Carlos Sanchez introduces Agentic Sites and plans to show what the team is building with AI for hyper-personalized websites.

[0:23](https://www.youtube.com/watch?v=jebp4V0vh30&t=23s) [Title slide for a presentation by Carlos Sanchez at Adobe.](https://assets.ai.engineer/sites/aiecode2025/sha256/dfc13a8de974b5de8b48ab788618de949b5328d1716b6a914b3a9f55f203558a.webp)

AGENTIC SITES: BUILDING HYPER
PERSONALIZED WEBSITES

Adobe

Carlos Sanchez / csanchez.org / @csanchez.org

### Speaker introduction

Carlos Sanchez is a principal scientist on Adobe Experience Manager, a content management product, and a longtime contributor to Kubernetes, Jenkins, and Apache Maven.

[0:39](https://www.youtube.com/watch?v=jebp4V0vh30&t=39s) [Speaker introduction describing a principal scientist at Adobe Experience Manager and open-source contributions.](https://assets.ai.engineer/sites/aiecode2025/sha256/bea10a18fd8d59b0eecb4467a92ddb340a5f66a1eedd8f751fe588c86d8eecd9.webp)

Principal Scientist

Adobe Experience Manager

Long time OSS contributor at Kubernetes, Jenkins,
Apache Maven,...

### Agentic Sites: What and How

Carlos Sanchez introduces what agentic sites are and how they are built.

[1:00](https://www.youtube.com/watch?v=jebp4V0vh30&t=60s) [Section title: Agentic Sites: What and How.](https://assets.ai.engineer/sites/aiecode2025/sha256/274996b91269120678bf7df36bf489aefefa287fa601a9f345dc1b269d089635.webp)

AGENTIC SITES:
WHAT AND HOW

### Intent-Driven Site

An intent-driven site personalizes pages in real time using a visitor’s persona and queries, with the goal of increasing engagement or conversions.

[1:05](https://www.youtube.com/watch?v=jebp4V0vh30&t=65s) [Intent-driven sites adapt to user goals and personalize pages in real time based on personas and queries.](https://assets.ai.engineer/sites/aiecode2025/sha256/b33ae4219ca10e7118307ea62c64baee326bfdaada950612d091c14794cbd9ac.webp)

INTENT-DRIVEN SITE

Adapts to each user's goals

Pages personalize in real time based on user persona and queries

### The Stack

AEM Edge Delivery puts content at the edge, while a backend connects to LLM services. Sanchez uses Cerebras for fast inference and says he will show it today.

[1:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=108s) [AEM Edge Delivery and an AI backend power the experience, with several model providers or local models.](https://assets.ai.engineer/sites/aiecode2025/sha256/a773318e63cb0decfe15deda5fc66165799fe595e7f8029d5b8795deb3413824.webp)

THE STACK

AEM Edge Delivery + a backend service power the experience

Combines Edge Delivery Services with an AI backend

Cerebras, Bedrock, Cloudflare, Vertex, even local models

### The Engine

The engine customizes site blocks for each user persona and uses the whole site as a corpus to ground generated content.

[2:23](https://www.youtube.com/watch?v=jebp4V0vh30&t=143s) [Personalization uses rich content, custom blocks, and the whole site as a content corpus.](https://assets.ai.engineer/sites/aiecode2025/sha256/c0692fb6d0d45d6612fc0d8c0dac9c1d397769ece3fddbf123bbf539f2a942eb.webp)

THE ENGINE

Personalization leverages rich content and custom blocks

and the whole site as content corpus for deep personalization

### Solving Static Site Limitations

Sanchez proposes using AI to create site variations, reducing the need to author thousands manually while enabling multiple layers of personalization.

[3:07](https://www.youtube.com/watch?v=jebp4V0vh30&t=187s) [Slide describing engagement and authoring limitations and multi-layer personalization.](https://assets.ai.engineer/sites/aiecode2025/sha256/0b8655a5dcc72a03184480ca5bca56bb83db60e8255e9c9fdbc952193642502f.webp)

SOLVING STATIC SITE LIMITATIONS

One-size-fits-all limits engagement
Manual authoring can't scale
Multi-layer personalization unlocks relevance

### Personalization Paths

Sanchez outlines instant persona adaptation, customized search results, and recommendation pages based on what visitors appear to be looking for after browsing.

[3:32](https://www.youtube.com/watch?v=jebp4V0vh30&t=212s) [Slide listing persona adaptation, on-demand query generation, and recommendations.](https://assets.ai.engineer/sites/aiecode2025/sha256/105ad5e6fde05afe1d4a627c72351a8b935add2a84e60fdf4ef458c861bf57ca.webp)

PERSONALIZATION PATHS

Instant Persona Adaptation
On-Demand Query Generation
Proactive "For You" Recommendations

### For Marketers

Marketers define the personalization strategy in natural language, then use analytics to adapt it in a continuing loop.

[3:59](https://www.youtube.com/watch?v=jebp4V0vh30&t=239s) [Slide describing natural-language personalization strategy and analytics-driven evolution.](https://assets.ai.engineer/sites/aiecode2025/sha256/201c39488f2ff4ef0ce5be0c19c1b7ad79053fa73809d69dc3f73e528d1176e4.webp)

FOR MARKETERS

Defining the personalization strategy in natural language

Continuously evolving the personalization strategy based on analytics

### Dynamic Frontend with Custom Blocks

The frontend combines custom blocks into flexible layouts. Sanchez says AI updates the blocks in real time through Edge Delivery Services.

[4:29](https://www.youtube.com/watch?v=jebp4V0vh30&t=269s) [Slide describing flexible layouts and real-time updates with AEM Edge Delivery Services.](https://assets.ai.engineer/sites/aiecode2025/sha256/113a012e2430eab235b9d0790f96221a61840745d0832453ebfdbdaa3cac04e1.webp)

DYNAMIC FRONTEND WITH CUSTOM BLOCKS

AEM Edge Delivery Services enables flexible layouts and real-time updates

### AI-Powered Backend Integration

A backend service fronts different LLMs for personalization. The team evaluates prompts across models and providers for accuracy and speed, tailoring evaluation to each site.

[4:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=284s) [Slide describing an LLM service for personalization and site-specific model evaluation.](https://assets.ai.engineer/sites/aiecode2025/sha256/1f462b2c766868aac29d9fb38765f383090a169911cf2baa249ed9761ff1c7b7.webp)

AI-POWERED BACKEND INTEGRATION

A service fronting any LLM — Cerebras, Bedrock, Vertex, local,... — for deep personalization

Model evaluation customized to each site to validate performance and accuracy

### Prompt Evaluation with Promptfoo

Sanchez uses Promptfoo continuously to compare prompts across 50+ models and providers on accuracy and speed, with evaluations tailored to each site. He aims to keep site generation within 1–2 seconds.

[5:29](https://www.youtube.com/watch?v=jebp4V0vh30&t=329s) [Promptfoo slide describing evaluation across models and providers for accuracy and speed.](https://assets.ai.engineer/sites/aiecode2025/sha256/434a3276764f351e5c310187e34e1908604a889be4764a92a4c9a70aaed61444.webp)

promptfoo

Evaluating different prompts against 50+ models and providers

Accuracy and speed

Different sites may need different accuracy and speed requirements

### Recommender benchmark results

On 15 prompts for an example site, Cerebras Gemma and Bedrock Ministral both passed 15/15 tests. Gemma averaged 1.1 seconds versus 4.6 seconds for Ministral; Sanchez says the speed difference drove their choice of Cerebras.

[6:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=395s) [Benchmark dashboard showing a 100% overall pass rate, 402 passed tests, and model latencies.](https://assets.ai.engineer/sites/aiecode2025/sha256/25bec058c8c7671b2f62398a6969850dcc9e9e629a73fd1a06c5fe25d84830dd.webp)

SUMMARY
100% Overall pass rate
402 Passed
0 Failed
59 Models
6/26/2026, 5:56:02 AM

MODELS
MODEL | PASS RATE ▼ | TESTS | AVG LATENCY
cerebras/gemma-4-31b-trial | 100% | 15/15 | 1.1s
bedrock/ministral-8b | 100% | 15/15 | 4.6s
bedrock/nova-lite | 100% | 15/15 | 4.8s
gemini-flash-lite-latest | 100% | 15/15 | 5.2s
gemini-flash-latest | 100% | 15/15 | 5.8s
gemini-3.5-flash | 100% | 15/15 | 7.6s
bedrock/mistral-large-3 | 100% | 15/15 | 8.1s
cerebras/glm-4.7 | 100% | 15/15 | 8.1s
gemini-2.0-flash | 100% | 15/15 | 8.2s
bedrock/nemotron-super-120b | 100% | 15/15 | 9.2s
vertex/llama-3.3-70b | 100% | 15/15 | 9.3s
bedrock/llama-3.3-70b | 100% | 15/15 | 9.7s

Interface text:
Intent Classification
Block Reasoning
Recommender

### Model benchmark comparison

Gemini 3.5 Flash passes 15/15 tests in 7.6s; Nemotron Nano passes 12/15 in 5.6s. Sanchez says a less-than-perfect model may be good enough if it is fast enough for the use case.

[7:32](https://www.youtube.com/watch?v=jebp4V0vh30&t=452s) [Scrolled benchmark table comparing model pass rates, test counts, and latency, including results below 100%.](https://assets.ai.engineer/sites/aiecode2025/sha256/5da3a502fbe73f686a7c17bc8586eaa7caf8d64cf86dbda62c34b3cb0f6ce13a.webp)

gemini-3.5-flash  100%  15/15  7.6s
bedrock/mistral-large-3  100%  15/15  8.1s
cerebras/glm-4.7  100%  15/15  8.1s
gemini-2.0-flash  100%  15/15  8.2s
bedrock/nemotron-super-120b  100%  15/15  9.2s
vertex/llama-3.3-70b  100%  15/15  9.3s
bedrock/llama-3.3-70b  100%  15/15  9.7s
bedrock/nova-pro  100%  15/15  10.3s
gemini-2.5-flash  100%  15/15  12.0s
bedrock/claude-sonnet-4  100%  15/15  13.6s
vertex/diffusion-gemma-26b  100%  15/15  19.3s
bedrock/deepseek-v3.2  100%  15/15  30.0s
cerebras/gpt-oss-120b  100%  15/15  33.1s
bedrock/claude-opus-4  100%  15/15  60.1s
bedrock/claude-haiku-4.5  93%  14/15  7.8s
  1 failing test
  14 passing tests
bedrock/claude-3-5-haiku  93%  14/15  19.3s
  1 failing test
  14 passing tests
bedrock/qwen3-32b  87%  13/15  8.4s
  2 failing tests
  13 passing tests
bedrock-eu/nemotron-nano-30b  80%  12/15  5.6s
  3 failing tests
  12 passing tests
gemini-2.5-flash-lite (production)  80%  12/15  9.4s
  3 failing tests
  12 passing tests
73%

Interface text:
Chrome  File  Edit  View  History  Bookmarks  Profiles  Tab  Window  Help
Wed Jul 1 16:26
demo  Agentic Sites:  Audience of C[illegible]  Precision Bre[illegible]  One perfect c[illegible]  One perfect c[illegible]  Admin  Arco Model B[illegible]  arco-googlet[illegible]  of1 Labs – De[illegible]  Cerebras Clou[illegible]  Ask Gemini
aem-growth-adoption.github.io/aem-growth-arco-benchmark/

### Real-time personalization engine

Browsing activity and queries feed an LLM that personalizes the site. The example adapts the hero, product cards, blog feed, navigation and calls to action to each persona.

[8:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=494s) [Slide explaining how browsing and queries adapt page content, calls to action, and navigation to user personas.](https://assets.ai.engineer/sites/aiecode2025/sha256/61b14f006b44653bd9d65496e7012def939d9ad331053123c424a6b336f15e28.webp)

REAL-TIME PERSONALIZATION ENGINE

• Browsing and queries adapt content instantly to user personas
• Hero, product cards, blog feed, and navigation change based on persona
• Call-to-action and navigation links are tailored for each persona

### For You proactive generation

Browsing signals become a personalized query for a prefetched For You page. The page can be regenerated as context changes, though repeated LLM calls have cost implications.

[8:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=530s) [Slide describing how browsing signals become a personalized query and a prefetched page that updates as context changes.](https://assets.ai.engineer/sites/aiecode2025/sha256/d0b201729d59adaed073bd467daff2983a712c36d1125e7965cfaafc70d6d839.webp)

"FOR YOU" PROACTIVE GENERATION

• Passive Browsing Signals: Tracks page visits, scrolls, quiz answers, and product interactions
• Transforms user activity into a personalized natural-language query
• Prefetched page appears as a navigation link for immediate viewing
• Re-triggers when user context changes, ensuring up-to-date personalization

### AI-Powered Page Creation

A query triggers a personalized page. AI uses its intent and journey stage to select blocks and follow-up suggestions; example intents include buying and seeking information.

[9:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=590s) [Queries are classified by intent and journey stage to select personalized page blocks and suggestions.](https://assets.ai.engineer/sites/aiecode2025/sha256/764fcb11a7cdfebbc0387b861a7e81530f6dbe3f738baafaa753e2324ef6b7b4.webp)

AI-POWERED PAGE CREATION

• Queries trigger dynamic, personalized pages
• Intent classification: Queries are mapped to distinct intent types and journey stages
• AI chooses relevant blocks and follow-up suggestions based on user needs

### Intent-Driven Block Selection

AI assembles and sequences frontend blocks around each user’s goals and journey stage. The palette includes products, guides, interactive tools, media, and utilities.

[10:28](https://www.youtube.com/watch?v=jebp4V0vh30&t=628s) [Personalized pages assemble and sequence blocks such as products, guides, tools, media, and utilities.](https://assets.ai.engineer/sites/aiecode2025/sha256/8ae26840555d9db2fed27f3017d060ad7808a66e09c259cc463aa52af49b61ac.webp)

INTENT-DRIVEN BLOCK SELECTION

• AI pages from a rich block palette to match user intent
• Frontend blocks enable dynamic, personalized assembly
• Block sequence adapts to each user's goals and journey stage
• Includes products, guides, interactive tools, media, and utilities

### Content Corpus: Rich Grounding

The example grounds LLM responses in product content, guides and other site material. It supports product comparisons and pages tailored to different users.

[11:12](https://www.youtube.com/watch?v=jebp4V0vh30&t=672s) [Products, guides, experiences, blog content, tools, and bundles ground tailored recommendations and comparisons.](https://assets.ai.engineer/sites/aiecode2025/sha256/ff91ed9e1ea546c6bafa12cdaf49fbc9c71f27d09cdaae38e4e755047ede94b3.webp)

CONTENT CORPUS: RICH GROUNDING

• Grounding on products, guides, experiences, blog, tools, and bundles
• Direct product comparisons help users make informed decisions
• Product pages are tailored for different user personas, ensuring relevant recommendations

### Personalization System Architecture

After 2+ page visits, browser signals form a “For You” query; hover can trigger pre-generation and caching. Cloud Run uses hybrid RAG and reasoning to stream HTML and evaluate page quality.

[11:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=695s) [A four-layer diagram connects browser signals and speculative generation to a backend, inference services, storage, and content delivery.](https://assets.ai.engineer/sites/aiecode2025/sha256/bd8ab1bd687a3798be2709591263d9fae09afbdc4a6c6df262117de323c2657a.webp)

Layer 1 — Browser (Client)

Authors publish in DA
AEM CDN
Page Browser
User browses

Signal collector
capture page visits
• scroll depth
• quiz answers
• interactions

sessionStorage
session context
• Query history
• browsing history
• Inferred profile

Persona cookie
personalize the hero banner
product cards

Browsing signals
‘For You’ query synthesized
xted + link revealed

signals
sessionStorage
on hover / predicted
speculative engine
streams SSE to Cloud Run

‘For You’ speculative system
Phase 1:
- After 2+ page visits, synthesizes personalized query from browsing context, reveals ‘For You’ nav link

Phase 2:
- Speculative engine tracks cursor, projects trajectory
- On hover or predicted hover (≥60% confidence, rate-limited)
- sends SSE to backend to pre-generate personalized
- caches result in sessionStorage

On user click, page renders instantly from sessionStorage result
instant render on nav click

For user-typed natural language queries, shown streams session context to Server-Sent Events (SSE)

Layer 2 — Cloud Run Backend (arco-recommender)
User query

1. Intent classification
Gemma 4 (served on Cerebras) classifies query using browsing context

2. Hybrid RAG
Keyword search against DA content in parallel with semantic vector search across Firestore (product, brew guide, FAQ embeddings)

3. Reasoning engine
Gemma 4 on Cerebras selects page blocks and plans layout

4. Page generation
Gemma 4 on Cerebras streams HTML back to browser as SSE chunks

5. Async analytics
Gemma 4 on Cerebras evaluates page quality across four dimensions, stores results in Firestore, and notifies browser

SSE to Cloud Run
SSE HTML stream
browser displays score

Layer 3 — Inference & Google Cloud Platform
Cloud Run

Cerebras
• Gemma 4 inference (intent classification, reasoning, page generation LLM calls)

Firestore
• vector embeddings (3 store analytics)
• analytics results

Vertex AI
text-embedding-005
semantic search

Secret Manager
• DA authentication token
• Cerebras API key

Analytics results
SSE event

Layer 4 — Adobe Experience Manager (AEM)

Document Authoring (da.live)
CMS for authors

Edge Delivery CDN
(aem.page / aem.live)
serves all pages

GitHub
code sync

### Audience of One Landing Page

Sanchez introduces Audience of One as the idea of personalizing marketing for each individual. The landing page frames each person as a market of one.

[12:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=755s) [Dark landing page presents an audience-of-one marketing message and a demo button.](https://assets.ai.engineer/sites/aiecode2025/sha256/dd5261d774bbb41b725a35ecd4027312095a0c354defab063b60cda701708eef.webp)

of1
HOME
WHY NOW
HOW IT WORKS
SEE IT LIVE

AUDIENCE OF ONE

The end of audiences.
The beginning of everyONE.

For decades, marketing meant shouting at segments. Categories. Cohorts. of1 makes every person a market of one.
The marketing nirvana is real — and it's simpler than you think.

WATCH DEMO

ONE

Interface text:
of1.live
Chrome  File  Edit  View  History  Bookmarks  Profiles  Tab  Window  Help
Ask Gemini

### Arco Coffee Homepage

Sanchez introduces a generated coffee-equipment site. Its homepage pairs the Arco Primo pitch with navigation to products and stories.

[12:51](https://www.youtube.com/watch?v=jebp4V0vh30&t=771s) [Coffee equipment homepage shows an espresso machine, product introduction, navigation, and search field.](https://assets.ai.engineer/sites/aiecode2025/sha256/9525ad967b9d927bd75e5279b4148b0ec0837cb1171522cb6bbed46ddcf7a7c1.webp)

One perfect cup.
Every morning.

The Arco Primo is built for people who want quality without complexity. Heat up in 90 seconds, pull a beautiful shot, and start your day.

Explore Machines
Find Your Brew Style

Interface text:
Arco
PRODUCTS
STORIES
EXPERIENCES
ABOUT
FOR YOU
Ask about coffee equipment...
WHITE PAPER

Chrome
File  Edit  View  History  Bookmarks  Profiles  Tab  Window  Help
main--arco--carlossg.aem.page
Ask Gemini
Wed Jul 1 16:32

### Water Chemistry and Espresso

Sanchez opens Water Chemistry and Espresso, a story on the generated coffee equipment site. The article pairs its explanation of water’s role in espresso with a photo of water being poured into a machine.

[12:59](https://www.youtube.com/watch?v=jebp4V0vh30&t=779s) [An espresso article with a photograph of water being poured into a machine.](https://assets.ai.engineer/sites/aiecode2025/sha256/356fb305c52b4da0923117d86913363c83e65c66c13f1f0e9843d8aa1493d532.webp)

Water Chemistry and Espresso: Why Your Water Matters

By Elena Rossi | February 10, 2026 | 7 min read

You have spent weeks dialing in your grind size, perfecting your dose, and timing every shot to the second. Yet something still tastes off. The culprit [illegible]

Interface text:
Arco
PRODUCTS
STORIES
EXPERIENCES
ABOUT
FOR YOU
Ask about coffee equipment...
WHITE PAPER

Chrome File Edit View History Bookmarks Profiles Tab Window Help
Wed Jul 1 16:32
main--arco--carlossg.aem.page/stories/water-chemistry-and-espresso

## Full transcript

[0:01](https://www.youtube.com/watch?v=jebp4V0vh30&t=1s) [music]

[0:13](https://www.youtube.com/watch?v=jebp4V0vh30&t=13s) >> Hello.

[0:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=14s) Thank you for coming. Um I'm going to

[0:17](https://www.youtube.com/watch?v=jebp4V0vh30&t=17s) talk to you about Agility Sites, how we

[0:20](https://www.youtube.com/watch?v=jebp4V0vh30&t=20s) call it as building hyper-personalized

[0:23](https://www.youtube.com/watch?v=jebp4V0vh30&t=23s) websites. I'm not going to just talk

[0:25](https://www.youtube.com/watch?v=jebp4V0vh30&t=25s) about it. I'm going to show you what

[0:27](https://www.youtube.com/watch?v=jebp4V0vh30&t=27s) we're building.

[0:28](https://www.youtube.com/watch?v=jebp4V0vh30&t=28s) Um I've been working on on this project

[0:31](https://www.youtube.com/watch?v=jebp4V0vh30&t=31s) for for a bit now, and

[0:34](https://www.youtube.com/watch?v=jebp4V0vh30&t=34s) we'll try to show you what is possible

[0:36](https://www.youtube.com/watch?v=jebp4V0vh30&t=36s) today with with AI.

[0:39](https://www.youtube.com/watch?v=jebp4V0vh30&t=39s) Uh I work at Adobe at a

[0:42](https://www.youtube.com/watch?v=jebp4V0vh30&t=42s) I'm a principal scientist at a product

[0:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=44s) that not many people know, Adobe

[0:45](https://www.youtube.com/watch?v=jebp4V0vh30&t=45s) Experience Manager, content management.

[0:47](https://www.youtube.com/watch?v=jebp4V0vh30&t=47s) We run a lot of

[0:49](https://www.youtube.com/watch?v=jebp4V0vh30&t=49s) uh

[0:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=50s) website properties for big brands, and

[0:53](https://www.youtube.com/watch?v=jebp4V0vh30&t=53s) my background is in in open source, uh

[0:56](https://www.youtube.com/watch?v=jebp4V0vh30&t=56s) contributing to to a lot of foundations

[0:58](https://www.youtube.com/watch?v=jebp4V0vh30&t=58s) and projects.

[1:01](https://www.youtube.com/watch?v=jebp4V0vh30&t=61s) What are Agility Sites, and how are we

[1:04](https://www.youtube.com/watch?v=jebp4V0vh30&t=64s) building this thing?

[1:06](https://www.youtube.com/watch?v=jebp4V0vh30&t=66s) So, we're looking for sites

[1:08](https://www.youtube.com/watch?v=jebp4V0vh30&t=68s) that are

[1:10](https://www.youtube.com/watch?v=jebp4V0vh30&t=70s) uh

[1:11](https://www.youtube.com/watch?v=jebp4V0vh30&t=71s) looking at the what intent the user

[1:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=74s) browsing

[1:15](https://www.youtube.com/watch?v=jebp4V0vh30&t=75s) uh has. What is the user doing? What is

[1:19](https://www.youtube.com/watch?v=jebp4V0vh30&t=79s) the user trying to achieve? And the end

[1:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=82s) goal is to personalize these pages for

[1:25](https://www.youtube.com/watch?v=jebp4V0vh30&t=85s) the for the current user browsing, so

[1:28](https://www.youtube.com/watch?v=jebp4V0vh30&t=88s) that eventually this

[1:31](https://www.youtube.com/watch?v=jebp4V0vh30&t=91s) uh drives uh higher engagement or uh

[1:34](https://www.youtube.com/watch?v=jebp4V0vh30&t=94s) conversions, whatever the marketing

[1:36](https://www.youtube.com/watch?v=jebp4V0vh30&t=96s) teams want to want to achieve.

[1:39](https://www.youtube.com/watch?v=jebp4V0vh30&t=99s) And these pages are personalized in real

[1:41](https://www.youtube.com/watch?v=jebp4V0vh30&t=101s) time based on the on the user that is

[1:45](https://www.youtube.com/watch?v=jebp4V0vh30&t=105s) uh accessing the site, and what is the

[1:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=108s) what is the user doing.

[1:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=110s) The stack we're using is AMH delivery.

[1:53](https://www.youtube.com/watch?v=jebp4V0vh30&t=113s) So, this is the part of the product we

[1:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=115s) we have,

[1:56](https://www.youtube.com/watch?v=jebp4V0vh30&t=116s) uh where all the content is on the edge

[1:59](https://www.youtube.com/watch?v=jebp4V0vh30&t=119s) and then we have back end service that

[2:02](https://www.youtube.com/watch?v=jebp4V0vh30&t=122s) powers this experience with

[2:05](https://www.youtube.com/watch?v=jebp4V0vh30&t=125s) different LLM providers LLM services

[2:09](https://www.youtube.com/watch?v=jebp4V0vh30&t=129s) we use Cerebras for fast inference or we

[2:13](https://www.youtube.com/watch?v=jebp4V0vh30&t=133s) can use also we tried bedrock and and a

[2:15](https://www.youtube.com/watch?v=jebp4V0vh30&t=135s) bunch of others. I'll be showing

[2:17](https://www.youtube.com/watch?v=jebp4V0vh30&t=137s) Cerebras today

[2:19](https://www.youtube.com/watch?v=jebp4V0vh30&t=139s) and you will see the reason why.

[2:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=142s) The

[2:23](https://www.youtube.com/watch?v=jebp4V0vh30&t=143s) the engine that is personalizing this

[2:27](https://www.youtube.com/watch?v=jebp4V0vh30&t=147s) this bits is

[2:30](https://www.youtube.com/watch?v=jebp4V0vh30&t=150s) using the rich content and blocks.

[2:33](https://www.youtube.com/watch?v=jebp4V0vh30&t=153s) So different blocks on the site are

[2:36](https://www.youtube.com/watch?v=jebp4V0vh30&t=156s) customized depending on on what the user

[2:38](https://www.youtube.com/watch?v=jebp4V0vh30&t=158s) persona is.

[2:40](https://www.youtube.com/watch?v=jebp4V0vh30&t=160s) We don't want the the whole site to be

[2:42](https://www.youtube.com/watch?v=jebp4V0vh30&t=162s) generated. I mean if you talk to

[2:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=164s) marketing people they they have a very

[2:46](https://www.youtube.com/watch?v=jebp4V0vh30&t=166s) strict brand guidelines. You don't want

[2:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=168s) to just

[2:49](https://www.youtube.com/watch?v=jebp4V0vh30&t=169s) come up with our have some

[2:51](https://www.youtube.com/watch?v=jebp4V0vh30&t=171s) hallucinations there. So the what is

[2:54](https://www.youtube.com/watch?v=jebp4V0vh30&t=174s) personalized is different sections of

[2:56](https://www.youtube.com/watch?v=jebp4V0vh30&t=176s) the site and we use the whole site as

[3:00](https://www.youtube.com/watch?v=jebp4V0vh30&t=180s) a corpus. We built a rack from the whole

[3:03](https://www.youtube.com/watch?v=jebp4V0vh30&t=183s) site. So what is generated is grounded

[3:06](https://www.youtube.com/watch?v=jebp4V0vh30&t=186s) on on the existing site.

[3:09](https://www.youtube.com/watch?v=jebp4V0vh30&t=189s) We tried to solve the problem where one

[3:12](https://www.youtube.com/watch?v=jebp4V0vh30&t=192s) size fits all. We want

[3:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=194s) hyper-personalized experiences. Also we

[3:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=196s) want to help

[3:18](https://www.youtube.com/watch?v=jebp4V0vh30&t=198s) our customers to do more automatic

[3:21](https://www.youtube.com/watch?v=jebp4V0vh30&t=201s) authoring. So not having to create

[3:24](https://www.youtube.com/watch?v=jebp4V0vh30&t=204s) thousands of different variations of the

[3:26](https://www.youtube.com/watch?v=jebp4V0vh30&t=206s) site but use AI for this

[3:29](https://www.youtube.com/watch?v=jebp4V0vh30&t=209s) and then do these multiple layers of of

[3:32](https://www.youtube.com/watch?v=jebp4V0vh30&t=212s) personalization.

[3:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=215s) Some examples of what we're doing or

[3:37](https://www.youtube.com/watch?v=jebp4V0vh30&t=217s) I'll show in the demo. It's

[3:40](https://www.youtube.com/watch?v=jebp4V0vh30&t=220s) instant persona adaptation, query

[3:43](https://www.youtube.com/watch?v=jebp4V0vh30&t=223s) generation when the user search for

[3:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=224s) something on the site, the page with the

[3:47](https://www.youtube.com/watch?v=jebp4V0vh30&t=227s) results is customized for them and also

[3:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=230s) uh, something like recommendations where

[3:52](https://www.youtube.com/watch?v=jebp4V0vh30&t=232s) after you browse the site for a period

[3:54](https://www.youtube.com/watch?v=jebp4V0vh30&t=234s) of time, we

[3:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=235s) we can create a page that recommends

[3:58](https://www.youtube.com/watch?v=jebp4V0vh30&t=238s) something based on on on what you are

[4:00](https://www.youtube.com/watch?v=jebp4V0vh30&t=240s) what we think you are looking for.

[4:02](https://www.youtube.com/watch?v=jebp4V0vh30&t=242s) For marketers, uh, they can define this

[4:05](https://www.youtube.com/watch?v=jebp4V0vh30&t=245s) strategy on natural language, and they

[4:07](https://www.youtube.com/watch?v=jebp4V0vh30&t=247s) can use analytics to

[4:10](https://www.youtube.com/watch?v=jebp4V0vh30&t=250s) to drive the loop of personalization,

[4:13](https://www.youtube.com/watch?v=jebp4V0vh30&t=253s) and what is the end goal, and how this

[4:17](https://www.youtube.com/watch?v=jebp4V0vh30&t=257s) goes back again to change to adapt the

[4:19](https://www.youtube.com/watch?v=jebp4V0vh30&t=259s) personalization to improve that uh,

[4:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=262s) whole cycle. Everybody's talking about

[4:24](https://www.youtube.com/watch?v=jebp4V0vh30&t=264s) loops in this conference, so that's

[4:26](https://www.youtube.com/watch?v=jebp4V0vh30&t=266s) that's one of the loops there.

[4:30](https://www.youtube.com/watch?v=jebp4V0vh30&t=270s) How the architecture look like? So, it's

[4:32](https://www.youtube.com/watch?v=jebp4V0vh30&t=272s) a dynamic front end with some blocks,

[4:34](https://www.youtube.com/watch?v=jebp4V0vh30&t=274s) what I mentioned before, and with uh,

[4:37](https://www.youtube.com/watch?v=jebp4V0vh30&t=277s) edge delivery services is basically you

[4:39](https://www.youtube.com/watch?v=jebp4V0vh30&t=279s) compose these blocks, and uh, they are

[4:42](https://www.youtube.com/watch?v=jebp4V0vh30&t=282s) updated on on real time through with the

[4:46](https://www.youtube.com/watch?v=jebp4V0vh30&t=286s) AI.

[4:47](https://www.youtube.com/watch?v=jebp4V0vh30&t=287s) The back end,

[4:49](https://www.youtube.com/watch?v=jebp4V0vh30&t=289s) uh, we

[4:51](https://www.youtube.com/watch?v=jebp4V0vh30&t=291s) we do the um,

[4:53](https://www.youtube.com/watch?v=jebp4V0vh30&t=293s) evaluation of the models and the

[4:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=295s) providers,

[4:56](https://www.youtube.com/watch?v=jebp4V0vh30&t=296s) and one thing we realized is is that

[4:58](https://www.youtube.com/watch?v=jebp4V0vh30&t=298s) this is very dependent on the site. So,

[5:01](https://www.youtube.com/watch?v=jebp4V0vh30&t=301s) we have a bunch of prompts, and we look

[5:04](https://www.youtube.com/watch?v=jebp4V0vh30&t=304s) uh, we run it across a huge variety of

[5:06](https://www.youtube.com/watch?v=jebp4V0vh30&t=306s) uh, models and providers, and then we

[5:08](https://www.youtube.com/watch?v=jebp4V0vh30&t=308s) look at the accuracy, we look at the

[5:10](https://www.youtube.com/watch?v=jebp4V0vh30&t=310s) speed, but this is going to depend

[5:12](https://www.youtube.com/watch?v=jebp4V0vh30&t=312s) highly on what type of site, like how

[5:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=314s) big is the site, how I don't know, what

[5:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=316s) different

[5:19](https://www.youtube.com/watch?v=jebp4V0vh30&t=319s) what different um,

[5:21](https://www.youtube.com/watch?v=jebp4V0vh30&t=321s) area is the site targeting, what what

[5:24](https://www.youtube.com/watch?v=jebp4V0vh30&t=324s) type of commerce it is, and so on. So,

[5:26](https://www.youtube.com/watch?v=jebp4V0vh30&t=326s) we we run this this

[5:29](https://www.youtube.com/watch?v=jebp4V0vh30&t=329s) um, evaluation continuously. We use uh,

[5:32](https://www.youtube.com/watch?v=jebp4V0vh30&t=332s) Promptfoo. Uh, anybody heard about

[5:34](https://www.youtube.com/watch?v=jebp4V0vh30&t=334s) Promptfoo?

[5:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=335s) Okay, some people.

[5:37](https://www.youtube.com/watch?v=jebp4V0vh30&t=337s) So, Promptfoo allows you to evaluate

[5:39](https://www.youtube.com/watch?v=jebp4V0vh30&t=339s) models um, prompts against my multiple

[5:43](https://www.youtube.com/watch?v=jebp4V0vh30&t=343s) models, providers, and you can do local

[5:46](https://www.youtube.com/watch?v=jebp4V0vh30&t=346s) models and any of the

[5:49](https://www.youtube.com/watch?v=jebp4V0vh30&t=349s) a bunch of open AI compatible

[5:52](https://www.youtube.com/watch?v=jebp4V0vh30&t=352s) uh providers and and uh

[5:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=355s) a lot of them, basically.

[5:57](https://www.youtube.com/watch?v=jebp4V0vh30&t=357s) We look for two things. Why? Accuracy.

[6:01](https://www.youtube.com/watch?v=jebp4V0vh30&t=361s) That's that's typically what people look

[6:03](https://www.youtube.com/watch?v=jebp4V0vh30&t=363s) for, but also we want the speed because

[6:05](https://www.youtube.com/watch?v=jebp4V0vh30&t=365s) we don't want the site generation to

[6:08](https://www.youtube.com/watch?v=jebp4V0vh30&t=368s) take

[6:09](https://www.youtube.com/watch?v=jebp4V0vh30&t=369s) more than 1 or 2 seconds,

[6:12](https://www.youtube.com/watch?v=jebp4V0vh30&t=372s) right? Because people uh

[6:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=374s) this is already

[6:17](https://www.youtube.com/watch?v=jebp4V0vh30&t=377s) uh proven that people want the the

[6:19](https://www.youtube.com/watch?v=jebp4V0vh30&t=379s) faster the site, the more conversions it

[6:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=382s) it generates or the the better the

[6:23](https://www.youtube.com/watch?v=jebp4V0vh30&t=383s) experience it is for the user.

[6:26](https://www.youtube.com/watch?v=jebp4V0vh30&t=386s) Um yeah, what I mentioned is different

[6:27](https://www.youtube.com/watch?v=jebp4V0vh30&t=387s) sites may have different requirements.

[6:29](https://www.youtube.com/watch?v=jebp4V0vh30&t=389s) Uh so, you may have to run this uh

[6:32](https://www.youtube.com/watch?v=jebp4V0vh30&t=392s) evaluation of models depending on the

[6:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=395s) site.

[6:37](https://www.youtube.com/watch?v=jebp4V0vh30&t=397s) This is a an ex

[6:39](https://www.youtube.com/watch?v=jebp4V0vh30&t=399s) uh we we secured this

[6:41](https://www.youtube.com/watch?v=jebp4V0vh30&t=401s) some of these queries, so we have a 15

[6:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=404s) prompts for this example site.

[6:46](https://www.youtube.com/watch?v=jebp4V0vh30&t=406s) Um we have uh

[6:49](https://www.youtube.com/watch?v=jebp4V0vh30&t=409s) at the top you can see with Cerebras on

[6:51](https://www.youtube.com/watch?v=jebp4V0vh30&t=411s) the Gemma 4 model that was announced

[6:53](https://www.youtube.com/watch?v=jebp4V0vh30&t=413s) last last week,

[6:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=415s) we can get an average latency of 1.1

[6:58](https://www.youtube.com/watch?v=jebp4V0vh30&t=418s) seconds generating a page.

[7:00](https://www.youtube.com/watch?v=jebp4V0vh30&t=420s) You you can compare that to the second

[7:02](https://www.youtube.com/watch?v=jebp4V0vh30&t=422s) one, which is 4.6 seconds, right? So,

[7:06](https://www.youtube.com/watch?v=jebp4V0vh30&t=426s) the difference is huge.

[7:09](https://www.youtube.com/watch?v=jebp4V0vh30&t=429s) And that's why uh we use Cerebras for

[7:11](https://www.youtube.com/watch?v=jebp4V0vh30&t=431s) for this use case.

[7:13](https://www.youtube.com/watch?v=jebp4V0vh30&t=433s) And uh

[7:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=436s) you can see that different providers,

[7:18](https://www.youtube.com/watch?v=jebp4V0vh30&t=438s) different models have different

[7:21](https://www.youtube.com/watch?v=jebp4V0vh30&t=441s) um

[7:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=442s) different speeds. And here is uh let me

[7:25](https://www.youtube.com/watch?v=jebp4V0vh30&t=445s) I can show you the whole

[7:28](https://www.youtube.com/watch?v=jebp4V0vh30&t=448s) thing here. Not this one, this one,

[7:30](https://www.youtube.com/watch?v=jebp4V0vh30&t=450s) right? So, at the at the bottom we have

[7:33](https://www.youtube.com/watch?v=jebp4V0vh30&t=453s) other other tasks.

[7:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=455s) Sometimes uh maybe some of them may be

[7:38](https://www.youtube.com/watch?v=jebp4V0vh30&t=458s) good.

[7:39](https://www.youtube.com/watch?v=jebp4V0vh30&t=459s) They don't need to be perfect, but

[7:41](https://www.youtube.com/watch?v=jebp4V0vh30&t=461s) they're good enough if they're fast

[7:42](https://www.youtube.com/watch?v=jebp4V0vh30&t=462s) enough. So, that's going to be the the

[7:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=464s) kind of

[7:45](https://www.youtube.com/watch?v=jebp4V0vh30&t=465s) decisions that you need to make on

[7:47](https://www.youtube.com/watch?v=jebp4V0vh30&t=467s) whether the model is good enough for

[7:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=470s) your use case or not.

[7:53](https://www.youtube.com/watch?v=jebp4V0vh30&t=473s) Yeah, we're looking Yeah, average 1.1

[7:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=475s) seconds. And then the the next ones are

[7:58](https://www.youtube.com/watch?v=jebp4V0vh30&t=478s) going from 4 seconds higher.

[8:01](https://www.youtube.com/watch?v=jebp4V0vh30&t=481s) And you don't need a huge LLM to do this

[8:03](https://www.youtube.com/watch?v=jebp4V0vh30&t=483s) sort of work because you are generating

[8:06](https://www.youtube.com/watch?v=jebp4V0vh30&t=486s) text, you are deciding where to put

[8:09](https://www.youtube.com/watch?v=jebp4V0vh30&t=489s) blocks and how to organize the website,

[8:12](https://www.youtube.com/watch?v=jebp4V0vh30&t=492s) you don't need a lots of information for

[8:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=494s) that.

[8:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=496s) So, this browsing and the queries

[8:20](https://www.youtube.com/watch?v=jebp4V0vh30&t=500s) uh is are being recorded. So, these are

[8:23](https://www.youtube.com/watch?v=jebp4V0vh30&t=503s) the metrics or the

[8:27](https://www.youtube.com/watch?v=jebp4V0vh30&t=507s) the the data we gather from the user,

[8:30](https://www.youtube.com/watch?v=jebp4V0vh30&t=510s) and this is fed into the LLM to

[8:33](https://www.youtube.com/watch?v=jebp4V0vh30&t=513s) personalize the site. And then in this

[8:36](https://www.youtube.com/watch?v=jebp4V0vh30&t=516s) example, we personalize the hero card,

[8:38](https://www.youtube.com/watch?v=jebp4V0vh30&t=518s) the products, the blog feeds, and and

[8:41](https://www.youtube.com/watch?v=jebp4V0vh30&t=521s) the navigation based based on the

[8:42](https://www.youtube.com/watch?v=jebp4V0vh30&t=522s) persona.

[8:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=524s) Also, what are some of the buttons like

[8:46](https://www.youtube.com/watch?v=jebp4V0vh30&t=526s) our call to action navigation, you can

[8:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=528s) also we can also personalize those.

[8:52](https://www.youtube.com/watch?v=jebp4V0vh30&t=532s) We

[8:53](https://www.youtube.com/watch?v=jebp4V0vh30&t=533s) we create and I'll show you the a for

[8:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=535s) you page, which is a recommendation.

[8:57](https://www.youtube.com/watch?v=jebp4V0vh30&t=537s) And this is a interesting one because

[9:00](https://www.youtube.com/watch?v=jebp4V0vh30&t=540s) this you could

[9:02](https://www.youtube.com/watch?v=jebp4V0vh30&t=542s) pre-generate, right? As the user browses

[9:05](https://www.youtube.com/watch?v=jebp4V0vh30&t=545s) your site, you gather these signals, and

[9:08](https://www.youtube.com/watch?v=jebp4V0vh30&t=548s) you could keep generating in this. So,

[9:10](https://www.youtube.com/watch?v=jebp4V0vh30&t=550s) in this case, you wouldn't need so such

[9:12](https://www.youtube.com/watch?v=jebp4V0vh30&t=552s) a big speed.

[9:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=554s) But

[9:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=556s) but that's interesting because it it

[9:17](https://www.youtube.com/watch?v=jebp4V0vh30&t=557s) would be if a user wanted to

[9:20](https://www.youtube.com/watch?v=jebp4V0vh30&t=560s) buy something, you could just say,

[9:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=562s) "Okay, for you, I will recommend these

[9:24](https://www.youtube.com/watch?v=jebp4V0vh30&t=564s) three products or or something like

[9:26](https://www.youtube.com/watch?v=jebp4V0vh30&t=566s) that."

[9:27](https://www.youtube.com/watch?v=jebp4V0vh30&t=567s) Um

[9:29](https://www.youtube.com/watch?v=jebp4V0vh30&t=569s) Yeah, and then they can see this

[9:31](https://www.youtube.com/watch?v=jebp4V0vh30&t=571s) recommendation, and if they go there,

[9:33](https://www.youtube.com/watch?v=jebp4V0vh30&t=573s) that that could be pre-fetched for them.

[9:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=575s) And obviously, you have to keep updating

[9:37](https://www.youtube.com/watch?v=jebp4V0vh30&t=577s) it as the user navigates around the site

[9:39](https://www.youtube.com/watch?v=jebp4V0vh30&t=579s) and and so on. So that that's also

[9:42](https://www.youtube.com/watch?v=jebp4V0vh30&t=582s) something to consider on the cost cost

[9:46](https://www.youtube.com/watch?v=jebp4V0vh30&t=586s) implications of doing multiple

[9:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=588s) generations, multiple LLM calls.

[9:54](https://www.youtube.com/watch?v=jebp4V0vh30&t=594s) When

[9:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=595s) when the user runs a query, dynamic

[9:58](https://www.youtube.com/watch?v=jebp4V0vh30&t=598s) personalized page is shown to them.

[10:00](https://www.youtube.com/watch?v=jebp4V0vh30&t=600s) When the

[10:02](https://www.youtube.com/watch?v=jebp4V0vh30&t=602s) these queries are also grouped into

[10:05](https://www.youtube.com/watch?v=jebp4V0vh30&t=605s) personas or intent types. So what is

[10:08](https://www.youtube.com/watch?v=jebp4V0vh30&t=608s) this guy what is this guy trying to do

[10:11](https://www.youtube.com/watch?v=jebp4V0vh30&t=611s) in the site? Is trying to buy something?

[10:13](https://www.youtube.com/watch?v=jebp4V0vh30&t=613s) Is trying to just get information? So

[10:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=616s) you can get marketers to decide what

[10:20](https://www.youtube.com/watch?v=jebp4V0vh30&t=620s) type of groups, how many groups you want

[10:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=622s) to have, how you want to deal with with

[10:23](https://www.youtube.com/watch?v=jebp4V0vh30&t=623s) customers. And the AI will choose the

[10:26](https://www.youtube.com/watch?v=jebp4V0vh30&t=626s) the blocks and the suggestions for for

[10:28](https://www.youtube.com/watch?v=jebp4V0vh30&t=628s) those groups of people.

[10:31](https://www.youtube.com/watch?v=jebp4V0vh30&t=631s) Um

[10:33](https://www.youtube.com/watch?v=jebp4V0vh30&t=633s) And we can adopt yes, the the different

[10:36](https://www.youtube.com/watch?v=jebp4V0vh30&t=636s) blocks, the the the sequence of the

[10:39](https://www.youtube.com/watch?v=jebp4V0vh30&t=639s) blocks and

[10:41](https://www.youtube.com/watch?v=jebp4V0vh30&t=641s) media. You could also do media. One of

[10:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=644s) the things we consider is

[10:46](https://www.youtube.com/watch?v=jebp4V0vh30&t=646s) there was some a model announced

[10:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=648s) today or yesterday the

[10:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=650s) the nano banana light. So you could even

[10:53](https://www.youtube.com/watch?v=jebp4V0vh30&t=653s) generate images

[10:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=655s) very fast on the fly.

[10:58](https://www.youtube.com/watch?v=jebp4V0vh30&t=658s) Obviously not as fast as text, but

[11:00](https://www.youtube.com/watch?v=jebp4V0vh30&t=660s) that's also something that would be

[11:03](https://www.youtube.com/watch?v=jebp4V0vh30&t=663s) I don't I don't know if it's that

[11:04](https://www.youtube.com/watch?v=jebp4V0vh30&t=664s) something like marketing people would

[11:06](https://www.youtube.com/watch?v=jebp4V0vh30&t=666s) want to have generated images. That

[11:09](https://www.youtube.com/watch?v=jebp4V0vh30&t=669s) depends on on the quality a lot if it's

[11:11](https://www.youtube.com/watch?v=jebp4V0vh30&t=671s) on brand.

[11:13](https://www.youtube.com/watch?v=jebp4V0vh30&t=673s) And

[11:15](https://www.youtube.com/watch?v=jebp4V0vh30&t=675s) the site in this example we have a a

[11:18](https://www.youtube.com/watch?v=jebp4V0vh30&t=678s) product site

[11:20](https://www.youtube.com/watch?v=jebp4V0vh30&t=680s) and then we have guides, experiences,

[11:21](https://www.youtube.com/watch?v=jebp4V0vh30&t=681s) blocks and the whole response of the LLM

[11:24](https://www.youtube.com/watch?v=jebp4V0vh30&t=684s) is grounded there. And

[11:27](https://www.youtube.com/watch?v=jebp4V0vh30&t=687s) there's comparisons. We can do

[11:28](https://www.youtube.com/watch?v=jebp4V0vh30&t=688s) comparisons between products that are

[11:30](https://www.youtube.com/watch?v=jebp4V0vh30&t=690s) tailor and the product pages can be

[11:32](https://www.youtube.com/watch?v=jebp4V0vh30&t=692s) tailored for the for the user.

[11:36](https://www.youtube.com/watch?v=jebp4V0vh30&t=696s) Okay, this is this is a bit of the

[11:38](https://www.youtube.com/watch?v=jebp4V0vh30&t=698s) stack.

[11:39](https://www.youtube.com/watch?v=jebp4V0vh30&t=699s) Um not going to spend too much time

[11:42](https://www.youtube.com/watch?v=jebp4V0vh30&t=702s) here, but the browser you have some

[11:45](https://www.youtube.com/watch?v=jebp4V0vh30&t=705s) layers. You have the browser where the

[11:47](https://www.youtube.com/watch?v=jebp4V0vh30&t=707s) signals get

[11:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=710s) get

[11:51](https://www.youtube.com/watch?v=jebp4V0vh30&t=711s) uh

[11:52](https://www.youtube.com/watch?v=jebp4V0vh30&t=712s) from I got I got I got so

[11:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=715s) I got from the from the user and then we

[11:58](https://www.youtube.com/watch?v=jebp4V0vh30&t=718s) have the back end. Uh

[12:00](https://www.youtube.com/watch?v=jebp4V0vh30&t=720s) we can have the back end. We run this

[12:01](https://www.youtube.com/watch?v=jebp4V0vh30&t=721s) some of these in in Google. Some of

[12:03](https://www.youtube.com/watch?v=jebp4V0vh30&t=723s) these are in our Cloudflare. So, the

[12:06](https://www.youtube.com/watch?v=jebp4V0vh30&t=726s) back end is basically just calling the

[12:07](https://www.youtube.com/watch?v=jebp4V0vh30&t=727s) LLM and doing some reasoning using the

[12:11](https://www.youtube.com/watch?v=jebp4V0vh30&t=731s) rack that is built on on the site to do

[12:13](https://www.youtube.com/watch?v=jebp4V0vh30&t=733s) the generation.

[12:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=736s) And you have obviously you have to have

[12:18](https://www.youtube.com/watch?v=jebp4V0vh30&t=738s) the vector database, the inference uh

[12:21](https://www.youtube.com/watch?v=jebp4V0vh30&t=741s) machinery and uh

[12:24](https://www.youtube.com/watch?v=jebp4V0vh30&t=744s) that obvious business manager is doing

[12:25](https://www.youtube.com/watch?v=jebp4V0vh30&t=745s) the serving the

[12:28](https://www.youtube.com/watch?v=jebp4V0vh30&t=748s) the

[12:29](https://www.youtube.com/watch?v=jebp4V0vh30&t=749s) at the edge is serving the the pages and

[12:31](https://www.youtube.com/watch?v=jebp4V0vh30&t=751s) the static content.

[12:33](https://www.youtube.com/watch?v=jebp4V0vh30&t=753s) So, let me show you because I think this

[12:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=755s) is uh so, we call this uh audience of

[12:38](https://www.youtube.com/watch?v=jebp4V0vh30&t=758s) one

[12:39](https://www.youtube.com/watch?v=jebp4V0vh30&t=759s) because the idea of in marketing

[12:43](https://www.youtube.com/watch?v=jebp4V0vh30&t=763s) they they always dream on being able to

[12:45](https://www.youtube.com/watch?v=jebp4V0vh30&t=765s) personalize things for each individual.

[12:49](https://www.youtube.com/watch?v=jebp4V0vh30&t=769s) So, we call it yeah audience of one. So,

[12:51](https://www.youtube.com/watch?v=jebp4V0vh30&t=771s) I have this this site. Uh this is a site

[12:54](https://www.youtube.com/watch?v=jebp4V0vh30&t=774s) that is absolutely generated

[12:56](https://www.youtube.com/watch?v=jebp4V0vh30&t=776s) uh example site. It's a coffee

[12:59](https://www.youtube.com/watch?v=jebp4V0vh30&t=779s) uh machinery. So, I can go and and read

[13:02](https://www.youtube.com/watch?v=jebp4V0vh30&t=782s) some stories

[13:04](https://www.youtube.com/watch?v=jebp4V0vh30&t=784s) and I can go and look at some products.

[13:08](https://www.youtube.com/watch?v=jebp4V0vh30&t=788s) Let's go and look at this product.

[13:11](https://www.youtube.com/watch?v=jebp4V0vh30&t=791s) I can spend some time here.

[13:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=796s) Uh

[13:17](https://www.youtube.com/watch?v=jebp4V0vh30&t=797s) let's go and click

[13:19](https://www.youtube.com/watch?v=jebp4V0vh30&t=799s) here.

[13:20](https://www.youtube.com/watch?v=jebp4V0vh30&t=800s) Okay, so I'm I'm browsing around the

[13:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=802s) site and I have this debugging tool

[13:25](https://www.youtube.com/watch?v=jebp4V0vh30&t=805s) thing uh which

[13:28](https://www.youtube.com/watch?v=jebp4V0vh30&t=808s) Uh

[13:29](https://www.youtube.com/watch?v=jebp4V0vh30&t=809s) let me go here, I think.

[13:37](https://www.youtube.com/watch?v=jebp4V0vh30&t=817s) Let's see.

[13:40](https://www.youtube.com/watch?v=jebp4V0vh30&t=820s) So, down there is the signals that the

[13:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=824s) that the browsing

[13:46](https://www.youtube.com/watch?v=jebp4V0vh30&t=826s) is giving us. So, I don't know if you

[13:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=828s) can see it much because I cannot see it

[13:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=830s) much. The So,

[13:52](https://www.youtube.com/watch?v=jebp4V0vh30&t=832s) the user is bucketed into the exploring

[13:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=835s) category. We have the pages that have

[14:00](https://www.youtube.com/watch?v=jebp4V0vh30&t=840s) have visited, and then we have how much

[14:02](https://www.youtube.com/watch?v=jebp4V0vh30&t=842s) time is spending on each page. All of

[14:04](https://www.youtube.com/watch?v=jebp4V0vh30&t=844s) this data is now available for the LLM.

[14:09](https://www.youtube.com/watch?v=jebp4V0vh30&t=849s) So,

[14:11](https://www.youtube.com/watch?v=jebp4V0vh30&t=851s) if I go here, I already have a for you

[14:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=854s) page that was generated for me

[14:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=856s) and

[14:18](https://www.youtube.com/watch?v=jebp4V0vh30&t=858s) based on my browser.

[14:21](https://www.youtube.com/watch?v=jebp4V0vh30&t=861s) And you will not notice that it's

[14:24](https://www.youtube.com/watch?v=jebp4V0vh30&t=864s) slightly different than everything else,

[14:26](https://www.youtube.com/watch?v=jebp4V0vh30&t=866s) but if I go here and I run a query

[14:29](https://www.youtube.com/watch?v=jebp4V0vh30&t=869s) like I want I'm looking for a coffee

[14:33](https://www.youtube.com/watch?v=jebp4V0vh30&t=873s) machine to

[14:36](https://www.youtube.com/watch?v=jebp4V0vh30&t=876s) uh prepare

[14:38](https://www.youtube.com/watch?v=jebp4V0vh30&t=878s) coffee

[14:40](https://www.youtube.com/watch?v=jebp4V0vh30&t=880s) while camping.

[14:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=884s) The site is this was just generated for

[14:47](https://www.youtube.com/watch?v=jebp4V0vh30&t=887s) me.

[14:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=888s) And then you're going to see some things

[14:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=890s) like the text is customized. Camping

[14:52](https://www.youtube.com/watch?v=jebp4V0vh30&t=892s) shouldn't mean compromising on your

[14:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=895s) uh whatever routine.

[14:57](https://www.youtube.com/watch?v=jebp4V0vh30&t=897s) Uh the coffee tips for camping um

[15:00](https://www.youtube.com/watch?v=jebp4V0vh30&t=900s) machinery that are being recommended are

[15:02](https://www.youtube.com/watch?v=jebp4V0vh30&t=902s) coffee agile and um or the nano, which

[15:05](https://www.youtube.com/watch?v=jebp4V0vh30&t=905s) are

[15:07](https://www.youtube.com/watch?v=jebp4V0vh30&t=907s) good for

[15:09](https://www.youtube.com/watch?v=jebp4V0vh30&t=909s) for the

[15:11](https://www.youtube.com/watch?v=jebp4V0vh30&t=911s) for a camping trip, right?

[15:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=914s) So, you saw how fast this was.

[15:17](https://www.youtube.com/watch?v=jebp4V0vh30&t=917s) I'm going to run it here something

[15:19](https://www.youtube.com/watch?v=jebp4V0vh30&t=919s) similar that I had here and I can run it

[15:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=922s) on the debug mode here.

[15:25](https://www.youtube.com/watch?v=jebp4V0vh30&t=925s) And you will see, let's make this

[15:27](https://www.youtube.com/watch?v=jebp4V0vh30&t=927s) bigger.

[15:30](https://www.youtube.com/watch?v=jebp4V0vh30&t=930s) Total time 164 seconds to generate the

[15:33](https://www.youtube.com/watch?v=jebp4V0vh30&t=933s) page. So, this includes a round trip to

[15:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=935s) the LLM. This is using Cerebras Gemma 4.

[15:38](https://www.youtube.com/watch?v=jebp4V0vh30&t=938s) So, the the Gemma model from Google

[15:40](https://www.youtube.com/watch?v=jebp4V0vh30&t=940s) running on Cerebras on their

[15:43](https://www.youtube.com/watch?v=jebp4V0vh30&t=943s) very fast chips.

[15:45](https://www.youtube.com/watch?v=jebp4V0vh30&t=945s) Uh we get

[15:47](https://www.youtube.com/watch?v=jebp4V0vh30&t=947s) 2,300

[15:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=948s) tokens per second.

[15:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=950s) Which is not bad.

[15:52](https://www.youtube.com/watch?v=jebp4V0vh30&t=952s) I would say.

[15:53](https://www.youtube.com/watch?v=jebp4V0vh30&t=953s) >> [snorts]

[15:54](https://www.youtube.com/watch?v=jebp4V0vh30&t=954s) >> And if I run it again, uh probably

[15:56](https://www.youtube.com/watch?v=jebp4V0vh30&t=956s) something like that.

[15:58](https://www.youtube.com/watch?v=jebp4V0vh30&t=958s) Uh the LLM time is 1 second. And again,

[16:01](https://www.youtube.com/watch?v=jebp4V0vh30&t=961s) 2,200 tokens per second.

[16:04](https://www.youtube.com/watch?v=jebp4V0vh30&t=964s) This is something that we only dreamed

[16:06](https://www.youtube.com/watch?v=jebp4V0vh30&t=966s) about before.

[16:10](https://www.youtube.com/watch?v=jebp4V0vh30&t=970s) On the on this site example site, we

[16:13](https://www.youtube.com/watch?v=jebp4V0vh30&t=973s) have some other options.

[16:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=976s) Uh

[16:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=976s) so, because we we've been showing this

[16:18](https://www.youtube.com/watch?v=jebp4V0vh30&t=978s) to customers, so we have the

[16:20](https://www.youtube.com/watch?v=jebp4V0vh30&t=980s) the ability to change the different

[16:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=982s) models, temper temperature, tokens, and

[16:24](https://www.youtube.com/watch?v=jebp4V0vh30&t=984s) so on. And we can uh

[16:28](https://www.youtube.com/watch?v=jebp4V0vh30&t=988s) we can

[16:29](https://www.youtube.com/watch?v=jebp4V0vh30&t=989s) show uh and try the different models and

[16:31](https://www.youtube.com/watch?v=jebp4V0vh30&t=991s) see how they behave. Besides the

[16:33](https://www.youtube.com/watch?v=jebp4V0vh30&t=993s) automatic test with Prompt Full, then we

[16:36](https://www.youtube.com/watch?v=jebp4V0vh30&t=996s) can uh manually come and and click

[16:38](https://www.youtube.com/watch?v=jebp4V0vh30&t=998s) things and see and see how that how that

[16:40](https://www.youtube.com/watch?v=jebp4V0vh30&t=1000s) works.

[16:42](https://www.youtube.com/watch?v=jebp4V0vh30&t=1002s) And uh

[16:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=1004s) we also have

[16:46](https://www.youtube.com/watch?v=jebp4V0vh30&t=1006s) uh OfOneLabs.

[16:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=1008s) So, we have we build this tool that

[16:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=1010s) generates an agentic site for any site

[16:53](https://www.youtube.com/watch?v=jebp4V0vh30&t=1013s) we want. So, if somebody wants to have a

[16:57](https://www.youtube.com/watch?v=jebp4V0vh30&t=1017s) demo for a customer, come here

[17:00](https://www.youtube.com/watch?v=jebp4V0vh30&t=1020s) and enter the URL. In less than an hour,

[17:02](https://www.youtube.com/watch?v=jebp4V0vh30&t=1022s) you have an agentic site. I did this

[17:03](https://www.youtube.com/watch?v=jebp4V0vh30&t=1023s) last week with the AI engineering site.

[17:07](https://www.youtube.com/watch?v=jebp4V0vh30&t=1027s) And

[17:08](https://www.youtube.com/watch?v=jebp4V0vh30&t=1028s) I got this site that is just a search

[17:11](https://www.youtube.com/watch?v=jebp4V0vh30&t=1031s) box and a few things.

[17:13](https://www.youtube.com/watch?v=jebp4V0vh30&t=1033s) And

[17:15](https://www.youtube.com/watch?v=jebp4V0vh30&t=1035s) let me open it here, the full page. Not

[17:18](https://www.youtube.com/watch?v=jebp4V0vh30&t=1038s) this one. Yeah, okay.

[17:20](https://www.youtube.com/watch?v=jebp4V0vh30&t=1040s) So, I could say

[17:23](https://www.youtube.com/watch?v=jebp4V0vh30&t=1043s) Europe AI conferences.

[17:25](https://www.youtube.com/watch?v=jebp4V0vh30&t=1045s) So, these suggestions are also AI

[17:27](https://www.youtube.com/watch?v=jebp4V0vh30&t=1047s) generated. And I get a page that is

[17:32](https://www.youtube.com/watch?v=jebp4V0vh30&t=1052s) more focused on

[17:34](https://www.youtube.com/watch?v=jebp4V0vh30&t=1054s) It should be more focused on on the on

[17:37](https://www.youtube.com/watch?v=jebp4V0vh30&t=1057s) this European conferences.

[17:40](https://www.youtube.com/watch?v=jebp4V0vh30&t=1060s) If I go back, did I go

[17:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=1064s) I can search for anything the same way I

[17:47](https://www.youtube.com/watch?v=jebp4V0vh30&t=1067s) did with with the Arco. So, I as a

[17:49](https://www.youtube.com/watch?v=jebp4V0vh30&t=1069s) specific There was someone that was

[17:52](https://www.youtube.com/watch?v=jebp4V0vh30&t=1072s) generating a good comparison side to

[17:54](https://www.youtube.com/watch?v=jebp4V0vh30&t=1074s) side.

[17:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=1075s) Let me see if this one. Okay, here. This

[17:57](https://www.youtube.com/watch?v=jebp4V0vh30&t=1077s) one.

[17:58](https://www.youtube.com/watch?v=jebp4V0vh30&t=1078s) I went and this generated a page with

[18:02](https://www.youtube.com/watch?v=jebp4V0vh30&t=1082s) a pretty good comparison. If I'm looking

[18:04](https://www.youtube.com/watch?v=jebp4V0vh30&t=1084s) at two conferences and I need to decide,

[18:06](https://www.youtube.com/watch?v=jebp4V0vh30&t=1086s) if I figure out that the user wants to

[18:08](https://www.youtube.com/watch?v=jebp4V0vh30&t=1088s) do that, this is great because that

[18:10](https://www.youtube.com/watch?v=jebp4V0vh30&t=1090s) gives them a side-by-side comparison on

[18:12](https://www.youtube.com/watch?v=jebp4V0vh30&t=1092s) the fly.

[18:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=1094s) Now,

[18:17](https://www.youtube.com/watch?v=jebp4V0vh30&t=1097s) this this is I think this is cool

[18:19](https://www.youtube.com/watch?v=jebp4V0vh30&t=1099s) already, but then we have I have this

[18:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=1102s) idea that

[18:24](https://www.youtube.com/watch?v=jebp4V0vh30&t=1104s) probably the I'm

[18:27](https://www.youtube.com/watch?v=jebp4V0vh30&t=1107s) a bunch of people are we are talking

[18:28](https://www.youtube.com/watch?v=jebp4V0vh30&t=1108s) about is the web that is is the web the

[18:31](https://www.youtube.com/watch?v=jebp4V0vh30&t=1111s) future still and so on. Nobody knows.

[18:34](https://www.youtube.com/watch?v=jebp4V0vh30&t=1114s) But we can also do something with this

[18:38](https://www.youtube.com/watch?v=jebp4V0vh30&t=1118s) with this audience of one, this

[18:40](https://www.youtube.com/watch?v=jebp4V0vh30&t=1120s) generative sites. So, imagine you have

[18:43](https://www.youtube.com/watch?v=jebp4V0vh30&t=1123s) you have your personal assistant and you

[18:46](https://www.youtube.com/watch?v=jebp4V0vh30&t=1126s) ask a query through in this case through

[18:49](https://www.youtube.com/watch?v=jebp4V0vh30&t=1129s) Google and you say I want to buy I don't

[18:52](https://www.youtube.com/watch?v=jebp4V0vh30&t=1132s) remember what the query said. It was

[18:53](https://www.youtube.com/watch?v=jebp4V0vh30&t=1133s) something like I want to buy a machine

[18:57](https://www.youtube.com/watch?v=jebp4V0vh30&t=1137s) and I get this on my Google TV.

[18:59](https://www.youtube.com/watch?v=jebp4V0vh30&t=1139s) Right? So, this is absolutely

[19:01](https://www.youtube.com/watch?v=jebp4V0vh30&t=1141s) personalized to my query.

[19:04](https://www.youtube.com/watch?v=jebp4V0vh30&t=1144s) Okay? No, go back.

[19:08](https://www.youtube.com/watch?v=jebp4V0vh30&t=1148s) This is absolutely personalized to my

[19:09](https://www.youtube.com/watch?v=jebp4V0vh30&t=1149s) query. So, I'm there in my living room.

[19:11](https://www.youtube.com/watch?v=jebp4V0vh30&t=1151s) I don't need a phone, I don't need a

[19:13](https://www.youtube.com/watch?v=jebp4V0vh30&t=1153s) computer, I don't need anything, just my

[19:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=1156s) voice and something that will

[19:19](https://www.youtube.com/watch?v=jebp4V0vh30&t=1159s) kind of show me

[19:21](https://www.youtube.com/watch?v=jebp4V0vh30&t=1161s) something that is absolutely

[19:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=1162s) personalized to to me.

[19:29](https://www.youtube.com/watch?v=jebp4V0vh30&t=1169s) Okay, so that one.

[19:31](https://www.youtube.com/watch?v=jebp4V0vh30&t=1171s) So,

[19:33](https://www.youtube.com/watch?v=jebp4V0vh30&t=1173s) what I was trying to show and hopefully

[19:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=1175s) you remember from this session is that

[19:38](https://www.youtube.com/watch?v=jebp4V0vh30&t=1178s) this is now possible.

[19:40](https://www.youtube.com/watch?v=jebp4V0vh30&t=1180s) It's only going to get better from here

[19:42](https://www.youtube.com/watch?v=jebp4V0vh30&t=1182s) on. It's only going to get cheaper, it's

[19:44](https://www.youtube.com/watch?v=jebp4V0vh30&t=1184s) only going to get faster.

[19:45](https://www.youtube.com/watch?v=jebp4V0vh30&t=1185s) And you will uh be able to have uh

[19:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=1188s) huge personalization options for sites

[19:52](https://www.youtube.com/watch?v=jebp4V0vh30&t=1192s) and for other things.

[19:54](https://www.youtube.com/watch?v=jebp4V0vh30&t=1194s) And you can do this with intent driven.

[19:57](https://www.youtube.com/watch?v=jebp4V0vh30&t=1197s) So, what is the what is my user trying

[19:59](https://www.youtube.com/watch?v=jebp4V0vh30&t=1199s) to do? What does my user want to buy?

[20:02](https://www.youtube.com/watch?v=jebp4V0vh30&t=1202s) These sort of questions. And you can uh

[20:06](https://www.youtube.com/watch?v=jebp4V0vh30&t=1206s) assemble a page just for them.

[20:08](https://www.youtube.com/watch?v=jebp4V0vh30&t=1208s) And you can also do this with uh

[20:11](https://www.youtube.com/watch?v=jebp4V0vh30&t=1211s) multiple models and and eventually

[20:14](https://www.youtube.com/watch?v=jebp4V0vh30&t=1214s) it's just going to be

[20:16](https://www.youtube.com/watch?v=jebp4V0vh30&t=1216s) faster and faster, right?

[20:18](https://www.youtube.com/watch?v=jebp4V0vh30&t=1218s) So,

[20:19](https://www.youtube.com/watch?v=jebp4V0vh30&t=1219s) that's it. Um

[20:21](https://www.youtube.com/watch?v=jebp4V0vh30&t=1221s) thank you for coming and I hope you you

[20:22](https://www.youtube.com/watch?v=jebp4V0vh30&t=1222s) got the idea. Thanks.

[20:24](https://www.youtube.com/watch?v=jebp4V0vh30&t=1224s) >> [applause]
