Tetrafluoride

ZeroShot QC is visual quality inspection that needs no training data. Small factories still inspect parts by eye, because automated inspection usually needs thousands of labelled defect photos per product, and they don't have them. With ZeroShot QC, a quality engineer writes the inspection rules in plain English, adds one photo of a good part, and gets a PASS or FAIL verdict with the reason and a box drawn on each defect. It is built on Meta's models: - SAM 3.1 finds the part in the photo, so the judge looks at it up close. - Muse Spark judges the part against the rules and the good reference photo, explains its decision, and draws a box around each defect it finds. - Muse Image creates synthetic defect images. The app scores its own rules on them, and Muse Spark rewrites the rules from its mistakes. A new version is kept only if it scores better on real photos it never saw. Hard rules are checked in code, not left to the model. A broken rule, a missed defect or the wrong part always fails. On the real MVTec AD dataset (5 product types, 150 photos), it reaches F1 0.92 without any training, and it boxes 63% of real defects. Rule tuning raised metal nut F1 from 0.82 to 0.88 on held-out photos. The app also includes an evaluation dashboard, a mode that drafts an inspection profile from an SOP document, and a line mode that inspects every part from a camera feed. It runs with one docker compose up command, and the bundled samples work without an API key.