How this site works

Architecture

Stack

Why these choices

• No vector database or RAG for the site's AI chat — context-stuffing plus AI Gateway prompt caching, deliberately choosing not to add retrieval infrastructure this corpus's size does not need.
• evaluateFit never returns a hire/no-hire verdict — only strengths, ramp areas and interview questions, each traced back to a corpus source.
• A committed eval-artifact snapshot serves as the degraded-mode fallback, so the site never serves an uncurated model answer when a rate cap is hit.
• Rate limiting and Turnstile bot verification gate both the human chat UI and the MCP tool surface, with separate per-surface budgets.

Grounding eval

As of 2026-07-30— no run recorded yet

This snapshot is a schema-valid placeholder (model: STUB_EVAL_MODEL_PENDING_WU34) — the real grounding-eval run has not landed yet. It renders through the exact same EvalArtifact schema the real run will satisfy once it does, so nothing else on this page changes when it lands.

Pass rate

0% (0/0)

Cost

$0.00

Runtime

0s

Verified answers on file

0

Design decisions

This site runs on one platform end to end: SvelteKit on Cloudflare Workers, Durable Objects for chat sessions and rate limiting, Cloudflare AI Gateway in front of the model provider, Turnstile for bot defense, and PostHog (EU) for product analytics. There is no separate database — the career corpus is typed code committed to the repo, not rows in a table, and eval results are build artifacts committed alongside it rather than a live service to keep running. AWS and Vercel appear elsewhere on this site as work he has shipped on; this site itself is deliberately all-Cloudflare.

The chat and the fit-check tool run on no vector database and no retrieval-augmented generation. The entire career corpus — every venture, every fact this site is allowed to state — is small enough to fit inside a single model context, so it is stuffed into the prompt whole and cached at the AI Gateway rather than chunked, embedded and retrieved. Building retrieval infrastructure for content this size would have added a failure surface — re-embedding, index drift, retrieval misses — for a problem this content does not have. Knowing when not to reach for RAG is itself the judgment on display here, not an afterthought or a cost-cutting shortcut.

Nothing the AI says ships because the prose looked right in review. Every change touching what it can claim is checked against a committed question suite — deterministic checks on hard facts (names, dates, the never-claim list) plus an LLM-judged rubric for whether an answer is actually grounded in the corpus — and a run that falls under threshold blocks the change from going out at all. The pass rate, the model, and the run date shown on this page are an "as of" snapshot from the last recorded run, published as a committed artifact rather than a live feed: it is stated as of the date it ran, and it is not silently kept current between runs.

The AI Gateway has a daily spend cap, and per-IP rate limits sit in front of it. When the cap is exhausted, the surface does not go dark and it does not serve an unchecked answer either — it falls back to a committed snapshot of answers that already passed the eval gate above, plus the FAQ. A visitor who arrives after the daily budget is spent still gets something true and grounded, honestly labeled as the degraded mode it is, rather than a flagship feature that quietly stops working.

As of when this was written, running the whole site on this stack — Workers, Durable Objects, AI Gateway spend, Turnstile — comes out on the order of a few dollars a month, not a fixed figure pulled from a live billing dashboard. That number will drift as traffic and usage do; it is restated here from time to time as of a point in time, deliberately not wired to a live meter, for the same reason the eval results above are a snapshot rather than a feed: a page about rigor that quietly goes stale is worse than one that says plainly when it was last checked.