This site

A site arguing that one architect can run a venture with agent fleets had to prove that claim by being built that way, not just asserting it. Decisions, architecture, and governance here are his — what to build, what never to claim, when to serve a degraded answer instead of a wrong one. Execution, on SvelteKit and Cloudflare Workers, was largely agent-driven under his direction, the same operating model described elsewhere on the site.

The chat and fit tools run on no vector database and no retrieval-augmented generation at all. The entire corpus — every venture, every fact this site is allowed to state — fits inside one model context, so it is stuffed into the prompt and cached at the AI Gateway instead. Building retrieval infrastructure for content this size would have been the wrong call. Knowing when not to reach for RAG is itself the judgment on display, not an afterthought.

evaluateFit, the tool a visiting recruiter can point at a job description, never returns a hire or no-hire verdict about its own principal — only strengths tied to a corpus source, ramp areas, and interview questions. Judging its own builder’s fit for a role was never something to automate; restraint here is a deliberate design constraint, not a missing feature.

Every change touching what the AI can say is checked against a committed eval suite before it ships — a real gate, not a formality, that can block a deploy outright. That gate is the evidence behind the site’s bigger claim about agent-driven execution: the fleet writes the code, but nothing reaches a visitor until it clears a test it does not get to grade itself.

When the AI Gateway spend cap runs out, the site does not go dark or serve an unchecked answer. It falls back to a committed snapshot of eval-verified answers, plus the FAQ — a degraded mode, honestly labeled as such, rather than a dead box. A visitor arriving after the budget is spent still gets something grounded and true.

The one thing to change: the drift checker — the mechanism that keeps rendered pages honest against the corpus — was wired in only after several pages already existed, rather than from the first page. Retrofitting that discipline cost more than building it in from the start would have. The colophon lays out the rest of the stack and the current eval snapshot in full.