jobsift

jobsift ingests job listings in the shapes they actually arrive in — pasted text, saved HTML, PDFs, inconsistent formats from different boards — and runs an LLM pipeline that extracts structured fields such as salary, tech stack, seniority, location and visa language. Extraction runs inside a bounded agentic loop: low-confidence results are diagnosed and retried with a different strategy, and anything still uncertain is routed to human review rather than passed off as fact. Accuracy is designed to be measured, not assumed — a hand-labelled golden dataset of real listings scores every pipeline change on per-field accuracy, so each loop strategy is tested against a baseline instead of taken on faith.