Keep a large business directory accurate with a local-model fact-checking pipeline

The owner of a directory of tens of thousands of specialty retailers and ranges keeps listings current with a custom agent harness: models verify facts against sources, decide which records are the same business, and extract services and opening hours, with every answer checked against what the agent actually saw and the evidence saved for later audits by bigger models. When hosted models began refusing routine research for this niche, the work moved to three local GPU servers running an open model, which also run the company's CI. One box handles about 4,500 validated jobs an hour at 98/100 accuracy on their benchmark, at a fraction of what equivalent Mac hardware would cost. They chose the more accurate quantisation because a wrong fact is expensive to purge.