hire-me-mcp

This portfolio itself: an open-source MCP server, RAG-backed interview agent and Next.js site over one typed career dataset — built as the demonstrable, inspectable evidence of how I ship production software.

What it is

hire-me-mcp is the project you are reading right now: a public MCP server, an embedded RAG-backed interview agent, and this website, all serving one typed career dataset. My recent professional work at House Numbers is proprietary and cannot be shown, so I built this repo to be the thing a recruiter or engineer can actually open and judge — every architectural decision, test, and CI gate is public, in the same shape I ship professionally.

Coding patterns

• A Turborepo + pnpm monorepo with typed package boundaries: packages/career-data (the Zod-validated content corpus and its cross-entity lint engine), packages/core (domain services, search, the pgvector store and ingestion pipeline), packages/agent (the interview agent, its prompt contract and evals) and packages/connect-metadata (generated client-connection snippets, drift-checked in CI).
• The web app reads career content only through a server-only content layer — never by importing career-data directly — enforced by lint rules, so a page can't hardcode or fork the data it renders.
• A structural no-PII analytics scrubber: tool-call and chat analytics record event shapes, never free text, so usage stats exist without storing what anyone typed.
• Fail-closed auth on the private stats route, and security headers with a per-request-nonce, strict-dynamic Content-Security-Policy.

Testing patterns

A full pyramid, every layer a real CI gate:

• Unit/integration tests (Vitest) co-located with every source file, enforced test-first by repo hooks and a pre-commit runner.
• Playwright e2e against a production build, plus a second preview suite that runs against the real Vercel preview deployment of every PR — navigation, accessibility scans, responsive checks, and content-correctness specs that independently re-read the career dataset to catch any page that drifts from it.
• Protocol-level MCP integration tests driving the real @modelcontextprotocol/sdk client against a locally started production server, and an MCP smoke suite against the deployed endpoint.
• Retrieval evals — recall/precision/MRR over a committed golden-query dataset with hard thresholds — run as a required PR check on a disposable Neon branch, so a change that degrades retrieval quality cannot merge.
• Agent evals scoring groundedness, gap honesty and relevance on real model output, plus Lighthouse performance budgets and MCP/chat latency percentile budgets, all enforced in CI from committed config.

AI patterns

• A public MCP server exposing the career data as typed tools with citations back to source records, over Streamable HTTP.
• RAG on Neon pgvector with task-typed Gemini embeddings (RETRIEVAL_DOCUMENT at ingestion, RETRIEVAL_QUERY at search), incremental ingestion keyed on content hashes — an unchanged corpus re-runs with zero embedding calls — and automatic production re-indexing on every content merge, which is how the chat can answer questions about this very project.
• An embedded chat agent grounded in that retrieval layer, with honesty guardrails (known gaps are acknowledged, claims are cited) held in place by the eval gates above.
• Deliberate free-tier quota engineering: CI jobs that call Gemini share a concurrency group and per-run call caps, so the whole pipeline runs on free tiers without tripping rate limits.

Why it mattered

It is the one project where nothing has to be taken on faith. The claims this portfolio makes about how I work — test-first, typed boundaries, eval-gated AI, honest content — are not descriptions of past work; they are enforced, inspectable properties of this repository.