Yuliya Fedorchenko released job_search_tool, an MIT Python pipeline that pulls roles from company ATS APIs and scores each one against your experience with Claude

Reading a hundred job postings a week is the tax on a real job search, and most of them do not fit the person doing the reading. Yuliya Fedorchenko released job_search_tool on 19 September, an MIT-licensed Python pipeline that fetches roles directly from company applicant-tracking APIs, filters and deduplicates them, optionally scores fit against a user's own experience with Claude, and reviews the results in a local web board. The repository has 92 stars and 56 forks after five days on GitHub.

The pipeline has three parts and the first one is free. It reads a list of target companies and pulls open roles directly from each company's Greenhouse, Ashby, Workable, Lever or SmartRecruiters board. It then queries the Adzuna and Remotive aggregators for a wider sweep. A local SQLite database keeps a memory of every job it has ever seen, so a rerun only ever surfaces new postings, and a filter file drops anything whose title or job description does not match the user's stack.

What the AI evaluation does

The second part is the optional AI step, and it is where an Anthropic API key comes in. Each filtered role, together with the user's own experience profile, is sent to Claude. The model returns a fit score from zero to one hundred, a list of concrete gaps, transferable strengths and risk factors, and an apply, consider or skip recommendation. Scores are stored in the same SQLite file, so a job that has already been scored is never scored again and never re-billed.

Fedorchenko says the AI step costs money, and the tool has a dry-run mode that shows what would be evaluated without calling the API. Users can also cap the evaluation to a small number for the first pass, which is meant for checking quality and cost before running the full list.

Where the review happens

The third part is a small Next.js board that reads and writes the same SQLite database. It shows each candidate with its AI score and the model's notes, and lets the user mark a role as applied, in interview, rejected, skipped or silent. Notes stay in the local database, so a job seeker can track their pipeline in the same place they screen it.

The repository ships pre-configured for the author's own search: Canada, Python and JavaScript. Fedorchenko names three files a new user must edit before their first run, because a wrong location pattern or a wrong stack list will reject every real role. The personal experience file is gitignored by default, so the private profile never travels with the code.

Source

• Repository: fedorchenko-juli/job_search_tool