Rumi
- Status
- Live
- Domain
- AI
- Stack
- PythonClaudeGitHub Actions
- Role
- Creator
- Year
- 2026
- Links
- Case study
An autonomous job-intelligence agent that reduces roughly 4,600 daily postings from 77 company boards to a short, evidence-backed shortlist.
Problem
For an engineer in Zimbabwe, most "remote" postings are noise: remote within the US or the EU, stated in paragraph six. Reading everything by hand cost hours every morning and still missed things.
Constraints
"Remote" hides two separate questions: can this employer hire someone where I live, and where do I sit? A residence visa is not a work permit, so keyword filters alone can't answer either one.
Architecture
collect → normalize → filter → judge → contacts → render → send
4,600 ~3,300 ~400 Haiku, GitHub, HTML, 07:00
postings unique survivors then Opus LinkedIn text GmailTen sources and 77 Greenhouse, Lever, and Ashby boards feed one pipeline. It runs on a GitHub Actions schedule and commits its own history back, so there is no server to keep alive.
Decisions & trade-offs
- Deterministic work first, the model last. Stale, region-locked, or wrong-discipline postings are dropped before any model reads them. Every posting a filter rejects for free is one I never pay to judge.
- Two models. Claude Haiku triages survivors in batches; Claude Opus scores the best 15 in depth. Swapping the deep model for Sonnet cuts cost roughly fivefold.
- Every verdict shows its basis. Eligibility notes quote the posting where it was explicit, so a wrong call is easy to spot and correct through a plain feedback file.
Impact
A typical morning ends with 4 to 10 roles, each with pay, odds, ghost-posting flags, and a person to contact. A run costs about $0.35 to $0.70, and the code carries 150+ tests with ruff and pyright clean.