AI agent · Personal project
Repo Cartographer
Point it at a public GitHub repository and ask a question. It reads the real code through the GitHub API — never cloning — and writes an onboarding guide that cites only the files it actually opened.
Objective
Answering “where does routing happen?” in an unfamiliar codebase takes judgement about what to read next, which a script cannot supply, and no context window holds a whole repository. The failure that matters is a guide confidently citing a file that does not exist, because a reader will trust it and go looking.
Solution
A LangGraph orchestrator sizes the repository up and fans out to as many as three explorers, one per top-level directory, each in its own context window. A doc-writer with no repository access builds the guide from their notes, and a link-checker with no model in it matches every cited path against the real file tree. Each guide states what it did not read.
My contribution
Sole author, built in nine phases with each definition of done fixed in advance and the measurement that proved it. The capability boundaries are pinned by tests that need no model, the GitHub tools layer is tested against the live API rather than mocks, and the one irreversible action is gated behind human approval and off by default.