Autonomous Job Board Finder
Give it a LinkedIn posting and it navigates to that company's own careers page, scoring each rendered page on eight calibrated signals to decide when it has arrived.
- Python
- Playwright
- Claude Haiku
- FastAPI
The problem
Job aggregators are a lossy copy of the truth. A posting on LinkedIn may be stale, may be reposted by a third-party recruiter, or may have already been filled — while the company's own careers page has the current listing and the real application form. The useful move is almost always to leave the aggregator and apply at the source.
Doing that by hand is tedious in a specific way: it is easy for a person and annoying at volume. Company career sites are inconsistent. Some are on Greenhouse or Lever, some are a custom React app behind a redirect, some bury the listings three clicks into a corporate marketing site. There is no single URL pattern to guess at.
Approach
The agent takes a LinkedIn posting URL and walks outward to find that company's own job board.
Pages are driven with Playwright, so the agent sees the DOM after JavaScript has run — a large fraction of career pages render their listings client-side and look empty to a plain HTTP fetch. Each rendered page is then scored by Claude Haiku against eight calibrated signals that together answer one question: is this the company's actual job board, or just a page that mentions jobs?
Scoring rather than pattern-matching is the important design choice. A URL allowlist would break on the first company that self-hosts, and a single yes/no LLM call is too brittle to trust. Multiple weighted signals let the agent express confidence and keep searching when no page clears the bar.
Result
Evaluated against 20 companies:
- 16 of 20 reached the correct job board on exact URL match
- 18 of 20 correct when verified by hand — the remaining two landed on a valid board that differed from the expected URL, for example a regional or localized variant of the same listing page
Running it
Deployed as a public FastAPI service that streams results as the agent walks, so you can watch it reason rather than staring at a spinner. It runs under a daily spend cap — the agent calls a paid API on every page it scores, and an unbounded public demo is an unbounded bill.