Job Search Agent System: Human-in-the-Loop Decision Support

Context: A job search is hundreds of small decisions a week: which posting to read, which to discard, what each CV should say. An AI agent reads and writes fast, but it can also invent a fact, forget a rule or attach the wrong file. This system pairs an agent with written rules, a library of verified facts and blocking checks.

Role: Sole designer and daily operator: process, rules, controls and data architecture. Much of the code was written by the agent under my direction and review. The private system is not published; the public repository holds its design and reference implementations written from scratch, with invented data.

Decisions: A human approval bound to a fingerprint of the exact text approved: if the text changes, the document is blocked. Gates block, never warn. Owner rules carry a date, so a changed rule reopens the discards it justified. One fact lives in one file: a per-posting CV only selects and orders fragments, it never writes a claim. The agent is evaluated against the owner's decisions in both directions.

Result: About 14,800 postings seen, 822 analyzed in writing, 77 documents sent, none without the owner approving its full text. 12 blocking checks per application. The owner authorized all 33 agent recommendations and reopened 5 of its 789 discards. Public repository with CI: tests, sanitization gate, a numbers check and a rendered-page test.

Technology Stack:

Claude Code (skills, hooks, memory, CLAUDE.md), Python (standard library + PyYAML), Bash, LaTeX (LuaLaTeX), poppler, Pandoc, Node.js + Playwright, HTML/CSS/JavaScript, YAML + JSON Schema, Ollama (local 3B model), GitHub Actions

About Team