Hypermode Launch

Kevin Van Gundy recalls three problems from his time at Vercel, when the company had an office above a pizzeria. It was losing to other JavaScript frameworks, losing badly to other hosting providers, and he was losing to a diet consisting exclusively of pepperoni pizza.

The turnaround, in his telling, came from developing the ability to iterate quickly. Without knowing the optimum strategy, the team could try more things, discover what the market wanted, and adapt its products accordingly. “Iteration is the compound interest of software.” Repeated experiments eventually revealed two developer preferences: adopt new technology incrementally, and defer architectural commitments until the application's behavior is better understood.

That experimentation stops when developers are afraid of getting things wrong. Generative AI adds a long chain of decisions: hardware, model, prompt, integration, monitoring, and improvement. The consequences are familiar—runaway bills, hallucinating chatbots, and projects that consume months without delivering value. A system that supports iteration must reduce the cost of mistakes, because mistakes will happen repeatedly.

Van Gundy identifies three ways to make those decisions easier to revisit:

• Model replacement: choosing the wrong model becomes less consequential when switching it out introduces little friction.
• Shared APIs: classical software and AI components become easier to integrate when they use the same interfaces.
• Inference tracing: following inference step by step gives developers visibility when changing prompts, strategies, or the mix of data supplied to a model.

Together, these mechanisms make changes easier to attempt and their effects easier to inspect.

Hypermode's proposed entry point is a runtime that integrates models and data into AI functions. The goal is to let developers add AI to an application without first acquiring specialized skills. Around that runtime, Hypermode supplies tools for rapid iteration and observation of functions in production. A team can begin with an incremental feature, develop its skills, and then reconsider the broader application as AI-native. The initial integration does not require making that larger architectural commitment upfront.

Even adding one model can involve unnecessary documentation work. Van Gundy uses temperature as the example: does a model expect a value from 0–2, 0–1, or 1–10? The developer needs that information while configuring the model. Hypermode's launch-era promise was typeahead in the developer's preferred editor, out of the box, with no SDKs or downloads. This describes the workflow presented at the launch, rather than current installation guidance. The editor popup on the slide illustrates bringing parameter documentation directly into that workflow.

For production, Hypermode offers strong defaults while also allowing teams to bring an existing stack. The concrete example is retrieval-augmented generation, or RAG: embedding an input and searching a vector store introduces additional service interactions into a previously simpler application request.

Van Gundy characterizes traditional RAG as requiring N+1 requests and says Hypermode consolidates that retrieval path into one request.

The consolidation concerns the embedding-and-search path; those operations still have to occur. Van Gundy reports a saving of a couple hundred milliseconds per request for this retrieval path. He supplies no benchmark configuration, baseline measurements, or latency percentile, so the magnitude cannot be assessed independently from the presentation.

Once the integration works, understanding its behavior remains difficult. Nondeterministic systems do not produce the same intuitive feedback as conventional software, and Van Gundy describes each model as having its own personality. Hypermode's comparison tools let developers examine different inferences, tunes, and models. The resulting dataset can then be exported for fine-tuning, connecting the work of inspecting outputs with a subsequent model-improvement step.

The closing question is practical: what will attendees tell their boss on Monday about what they learned? Van Gundy invites them to the after-lunch Hypermode workshop, with the team that built the platform, to practice making iteration velocity a core competency. The proposed applications are natural-language search, intelligent sorting of product data lists, outlier detection, and identifying bad actors.

The workshop promise was a demo attendees could be proud of and a plan to put something similar into production by the end of the following month. Van Gundy also offered $1,000 in Hypermode credits to help attendees get started. These were offers made at the 2024 event. The intended next step was to carry a concrete application beyond the workshop, with a production plan that made continued iteration possible.