Saw_the_bay_bridge_while_working_NICE

Catalog Harness helps retailers turn inconsistent supplier catalogs into evidence-backed product decisions. It preserves raw inputs, normalizes supplier fields, and uses an OpenAI model to produce structured categories and product uses with exact source citations. Conflicting evidence triggers abstention and human review. The project demonstrates a bounded harness-improvement experiment: revise the policy while keeping the model, taxonomy, inputs, and scoring fixed; select the candidate on validation; then evaluate on a locked synthetic holdout. Validation improved from 33/36 to 34/36 correct field decisions, with unsupported assertions dropping from four to zero. Final holdout results improved from 68/72 to 69/72, with unsupported assertions falling from two to one. A deterministic placement layer maps supported decisions into store-specific categories. MongoDB Atlas persists inputs and experiment results, with a Raindrop integration for model observability. A separate 50,000-record synthetic database benchmark verified processing integrity, duplicate-free reseeding, and indexed query efficiency without additional model calls. Live demo: https://catalog-harness.onrender.com/