Job Matcher AI

Recruiters spend time manually mapping a candidate profile to a role. I wanted to invert that process: let the portfolio run the analysis and surface fit, gaps, and evidence in seconds. The hard part is not calling an LLM: it is making the output structured, consistent, and visually scannable instead of free-form text.

A portfolio feature where hiring managers paste a job description and receive a fit report across seven dimensions: ATS keywords, responsibilities, seniority, culture, critical gaps, minor gaps, and differentiators. The model returns JSON only, no prose outside the schema, so the UI can render badges, a radar chart, and grouped evidence reliably on every run.

I designed the scoring vocabulary (Strong Match, Good Match, Partial Match, Gap), the dimension tabs versus radar view, and the integration with the structured project bank so each claim links to real case evidence. Usage limits per portfolio password are stored in Vercel KV to keep API cost predictable while the feature stays accessible.

React + TypeScript frontend. Claude API (claude-sonnet-4-6) with a strict system prompt and JSON schema. Vercel serverless endpoints. Vercel KV for per-password usage counters. Shared parsing and validation with the CV generator pipeline where narratives must stay separate from audit language.

What shipped in production on chemaamate.es/match.

A public demonstration of AI product design end-to-end: prompt architecture, structured output, visualization, access control, and copy that helps recruiters decide faster, with a live path to try the matcher when they have a portfolio password.