Job Scraper 02

AI scoring is optional, not required. With no key it drops to keyword matching, still ranks the jobs and still works. That's the whole pitch: a job tool that makes you sign up for a paid API before it will show you anything is a job tool you close.

Everything the dashboard does is also a REST endpoint and a CLI command, so it wires into your own automation. There's an optional n8n profile if you want a full automation engine, and one-click configs for Railway and Render. The default is still docker compose up on your own machine, with a SQLite file you keep.

Lite Mode had been scoring every job zero since it shipped. A regex bug meant no listing ever matched, however well it fit the CV. No error, nothing in the logs — the failure mode that makes a feature look present and be useless. Fixed, with regression tests covering match, no-match, word-boundary and score-cap.

Notifications were dead by default — the alert threshold was 70 on a scale that only goes to 10. Results weren't sorted by score, despite the README claiming high scorers bubble to the top. Two scrapes could race if triggered close together; the run lock is atomic now. The plain `Dockerfile` most platforms auto-detect started only the UI with no backend attached, so it was silently unusable.

CI now proves the app works, not that it compiles. It builds the wheel and checks nothing leaked out of the package, installs and runs the real job-scraper console script, and drives a genuine scrape against the live containers, asserting a real scored job comes back. Same lesson as every other project of mine: a green pipeline that only ran the type checker means nothing.

Packaged properly. The flat top-level db / app / scraper modules moved into a real job_scraper package, so installing it can't collide with someone else's db module in their environment.