Job Search Copilot

There are two ingestion pipelines. The stored one persists jobs from sources that allow it — Greenhouse and Lever. The ephemeral one runs at query time, caches results in memory with a short TTL, and never writes to disk. The split exists because one of the sources I wanted to use is a job aggregator whose terms of service don't allow storing scraped data. Splitting lets me use it legally. Both paths share the same fetcher, normalizer, embedding, and ranking code underneath.

I designed the system around a ~$5/month budget. There's no runtime meter enforcing it — the budget shows up in the weight choices. Scoring every job with an LLM was off the table, so the ranking pipeline leans on deterministic rules: skill coverage (25%), title match (15%), seniority distance (10%), location fit (10%). Those rules cost nothing to run. 60% of the final score comes from them, and only the remaining 40% semantic-similarity portion actually touches an embedding. That weight split is what keeps the matcher cheap enough to run continuously instead of in occasional batches.

The obvious next step is an auto-apply bot. But auto-applying only works if the matching underneath is good. Otherwise you're blasting applications at jobs you don't fit and wasting time on both ends. So I built the matching layer first and stopped there. It's useful on its own — I can point it at my resume any week and get a ranked list with explanations, no application automation needed. Auto-apply becomes a future module, not the whole premise.