Freelance Bidding Agent: Decomposing Retrieval Hurt, a Simple Budget Rule Helped

A student built GigBid, a multi-agent system that reads freelance job posts, decides whether to bid, and drafts proposals whose claims must trace to portfolio evidence, with a human approving each one. The author hypothesized that splitting a job post into separate requirements would improve retrieval, but measured results showed it performed markedly worse than a simpler hybrid search with reranking. An ablation traced the problem to merging per-skill rankings, which rewarded projects matching one requirement over those matching all of them. Separately, the bid-or-skip decision was failing on low-budget jobs that the system never compared against the freelancer's rate floor. A short deterministic rule fixing this raised bid-or-skip F1 from 0.89 to 0.98 without changing any model or prompt.