CareerForge: evidence-backed career assets
Built a local-first CLI that turns Git history and AI coding sessions into résumé bullets where every claim cites its evidence, and refuses the claims the evidence cannot carry.
Version 0.2.1 published on npm. The full test suite runs on Linux, macOS, and Windows on every commit, and 30 architecture decision records document the design.
The problem
AI résumé tools write confident claims about work they cannot see. When no evidence goes in, nothing a model writes can be traced back to what actually happened, and the person is left defending numbers in an interview that no record supports.
What I built and owned
It refuses to invent
A model may propose a role, a metric, or an outcome, but generation keeps only the assertions with supporting records. In the guided tour, three of four proposed claims are removed, and each one names the question whose answer would make it supportable: what was your role, what result can you quote, and what changed as a result.
AI is optional, and the build enforces it
Dependencies point inward only. The domain layer performs no I/O and has no knowledge that AI exists, so collection, storage, search, timelines, the provenance graph, and export work with no API key and no network.
Privacy is architecture, not a promise
There is no CareerForge server, account, or telemetry. A provider receives a policy decision containing the exact redacted bytes that were approved, never a raw payload, and the user can preview those bytes first. AI coding transcripts are classified restricted by default, and every evaluation is recorded with a hash of the payload rather than the payload itself.
Nothing exports until it is reviewed
Evidence is append-only: corrections supersede and deletions tombstone. The review gate sits in the export path, so scripts inherit it. Storage is SQLite, but the durable copy is a plain JSON export, and a round-trip test in CI proves the database can be rebuilt from it.
Screenshots
