Nov 9, 2026
An Agent That Reads 4,600 Job Postings So I Don't Have To
A daily job-hunting agent that cuts thousands of postings down to a short, honest shortlist, using cheap deterministic filters first and Claude only where judgment pays.
The problem
Job boards are built to show you as many postings as possible. For an engineer based in Zimbabwe, most of them are noise. "Remote" usually means remote within the US or the EU, and the posting only says so in paragraph six.
Reading everything by hand cost hours every morning and still missed things. I wanted one short list in my inbox each day, with the reasoning attached.
What I built
Rumi is a daily agent that collects postings from ten sources and 77 company hiring boards, throws away about 99% of them, has Claude judge what's left, and emails one digest at 7am.
collect → normalize → filter → judge → contacts → render → send
4,600 dedupe to ~400 Haiku GitHub HTML + Gmail
postings ~3,300 survivors then + text 07:00
Opus LinkedInEach stage narrows on a different axis:
- Collect. Greenhouse, Lever, and Ashby boards (exact and free, as JSON), plus aggregators, Hacker News "Who's Hiring", and local Zimbabwean sources.
- Normalize. About 4,600 raw postings dedupe to roughly 3,300 unique roles.
- Filter. Cheap, certain cuts happen first, before any model sees anything: stale postings, the wrong discipline, region-locked roles, roles that aren't remote. Roughly 400 survive.
- Judge. Claude Haiku triages survivors in batches, then Claude Opus scores the best 15 in depth.
- Contacts. For each role, a way to reach a human on the team through GitHub and LinkedIn.
- Send. One HTML email, with anything sent in the last 30 days suppressed so the digest never repeats itself.
A typical morning ends with 4 to 10 roles.
The hard part: who can actually hire you
The constraint that shaped everything is that "remote" has two separate questions hiding inside it:
- Can the employer hire someone where I live? Most remote roles can't, whatever the headline says.
- Where do I sit? A digital nomad visa grants residence, not work authorisation, so moving countries doesn't make a "must have EU work rights" role eligible.
Keyword filters catch the obvious cases. The model resolves the ambiguous one, where "Europe" might mean a timezone or a right to work. Every verdict carries a note saying what it was based on, quoting the posting where the posting was explicit.
What each role comes with
- Pay, either stated or estimated, with the estimate's basis shown
- Realistic odds, and whether the application is worth the time
- Ghost-posting and toxicity flags
- An eligibility note explaining the verdict
- A person to contact
Cost and reliability
- About $0.35 to $0.70 per run, or $10 to $20 a month. Switching the deep model to Sonnet cuts that roughly fivefold.
- Runs on a GitHub Actions schedule and commits its history back, so it survives without a server.
- 150+ tests, with ruff and pyright kept clean.
- Steered through a plain feedback file: when Rumi gets something wrong, I write the correction there and it changes tomorrow's scoring.
What I learned
Put the deterministic work first and the model last. Every posting a filter can reject for free is a posting you never pay a model to read. The model is only useful where the answer genuinely needs reading.