# ZeroDefect

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Monashini S](https://cerebralvalley.ai/u/Monashine)
- **GitHub:** https://github.com/Monashini/zerodefect
- **Demo video:** https://youtu.be/V-uBvXXySCs
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/923

ZeroDefect is AI quality inspection for small factories that needs only one good part and a phone camera. Small units in Chennai's industrial estates (Ambattur, Guindy, Sriperumbudur) still check bolts, nuts, bottles and PCBs by eye, so defects slip through and whole batches get rejected. Machine-vision systems cost lakhs of rupees and need thousands of labelled defect photos.

How it works: the owner shows ZeroDefect one golden sample. SAM 3.1 finds every part in the frame from a phrase like "metal nut". A zero-training patch-feature detector (frozen ResNet-18, PatchCore-style memory bank built from the golden sample) makes an anomaly heatmap for each part. Muse Spark acts as the QC inspector: it compares the golden sample, the part and the heatmap, and returns a structured verdict (defect type, location in plain words, severity, likely process cause, a spoken call-out and a grounded defect box). That box goes back to SAM 3.1 as a box prompt to get a pixel-accurate defect mask and the defect's size. When the detector and Muse Spark disagree, the part goes to human REVIEW. One tap on "Actually OK" adds a part to the golden set, and a Muse Spark tool-calling agent turns the inspection log into a shift report. Everything falls back to local methods if the network drops.

Results with 1 golden image and 0 defect images on MVTec AD: bottle image AUROC 0.993 / pixel AUROC 0.983; metal nut (scratch + colour defects) 0.931 / 0.951. In the full app with 4 extra OK parts for calibration: 63/63 defective bottles caught (1/20 good flagged) and 36/45 defective nuts caught (0/22 good flagged). About 0.15 s per part on a laptop CPU. Built during the hackathon.

## More from Meta Global AI Developer Hackathon

- [kimdaeheuk](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/920)
- [Open Indiana Law](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/921)
- [Eddy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/922)
- [DUKERS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/924)
- [UFAT](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/925)
- [atlas](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/926)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/923. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt