Shipped a production AI product: a Cloudflare Worker backend serving four switchable engines; a résumé-grounded streaming assistant with citations, a job-fit analyzer, five-language translation, a voice agent, and an MCP server exposing five tools to any AI client. Owned architecture, model selection, and risk controls end to end: explicit refusal policy, rate limiting, PII-masked usage logging, a paid-engine spend cap, and 100 Lighthouse accessibility behind a CI quality gate. Authored a rubric-based evaluation of four production AI engines — a fixed 20-question suite scored 0–5 across four criteria, published across thirteen immutable runs with full transcripts and the judge's own reasoning. Since September 2026 the site also ships the Candidate Card (averyresume.com/candidate): a fully visible, one-response instruction a recruiter pastes into their own AI after a job description; it fetches the versioned public profile and returns evidence-backed matches, material gaps, five interview questions and one reason not to hire — and stops rather than guesses if the profile cannot be fetched. Live at https://averyresume.com — AI clients can connect to the MCP server at https://ai.averyresume.com/mcp