RemoteStar
A matching engine that reads roles the way a recruiter does.
Overview
RemoteStar matches candidates to companies with real open roles. The engine has to reason like a recruiter: not just embedding similarity, but function-aware scoring against what each company is actually hiring for. I own it end to end, from the pipeline to the product surface.
Challenge
Naive matching is either too expensive or too wrong. Scoring every candidate against every company burns API budget; pure vector similarity confuses ‘knows React’ with ‘right for this role.’ And LLM calls in the request path were timing out in production.
Approach
LLM scoring layered on Pinecone semantic search, with an OpenAI company-size pre-filter that cut backfill costs to around $10. A weekly BD-signals pipeline processes ~16,000 jobs across ~2,000 companies with freshness filtering, deduplication, change detection, and automated staleness handling. Long-running LLM work moved to a background queue with bounded polling so nothing times out in the request path.
Outcome
The engine runs in production with function-aware scoring across real open roles. Around it I shipped MPC profiles, job citations, redacted CVs, and outreach tooling in Next.js/React, plus Apollo contact discovery with a SignalHire fallback and caching. I own CI/CD and production backfills; the team brought me back for a second stint.