# Vulcan

- **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:** [Aditya Arnav](https://cerebralvalley.ai/u/hydraofmalice)
- **GitHub:** https://github.com/Hydra-Of-Malice/defect_forge
- **Demo video:** https://www.youtube.com/watch?v=dqv-X4mfdy8&t=1s
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2583

Defect Forge measures where a visual inspector stops seeing defects, before a production line makes any.

The problem: a new line has no photos of bad parts, so nobody knows what its inspector will miss until defects ship. The smallest and rarest defects are the ones with the least data.

What it does: you give Defect Forge your QA spec and a few photos of good parts. No real defect images are needed.

1. Muse Spark reads the spec into a defect list and finds the part in each photo. SAM 3.1 masks it.
2. Muse Image paints each defect onto a crop of a good part. Our deterministic compositor keeps only the changed pixels, so everything outside the defect stays byte-identical.
3. SAM 3.1 cross-checks the defect mask, a Muse Spark critic rejects weak forgeries and rewrites the prompt, and eight deterministic validators gate every sample.
4. Plain OpenCV code fades and shrinks each accepted defect on a fixed severity ladder, and Muse Spark inspects every variant over three passes.
5. Defect Forge plots detection-threshold curves and exports a labeled COCO evaluation suite. A live Inspect screen checks a single photo and reports "uncertain" when no SAM mask backs up a claimed defect.

Because we place each defect ourselves, ground truth is known by construction. This is a measurement, not a model grading itself.

Results on hazelnuts: forged cuts were caught 73% of the time at full size (about 1.5 mm) and 0% at about 0.5 mm. Sub-millimetre holes were usually noticed but not pinpointed. On the real MVTec AD hazelnut test split, the spec-only inspector reached 94-97% recall across three passes with no false alarms on 40 good images; a separate set of spec-only passes had 1 false alarm in 120. Its weak spot is real white print, read as gloss. Each accepted, pixel-labeled defect cost about $0.02 and 25 seconds.

What we don't claim: the curves are measured on forged defects, and below 1.7 mm, the smallest real defect in the benchmark, they have no real-world check yet. Forged examples in the prompt did not beat real good photos, and they lowered detection of faint defects (72 to 59 of 141). We wrote the test plan first and publish these results, plus a claim we withdrew, in the repo.

Hosted demo: https://defect-forge-kappa.vercel.app (a replay of a recorded live run; live mode runs locally from the repo).

## More from Meta Global AI Developer Hackathon

- [Hack Singh](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2580)
- [George (GK) — TRUE IP / FME, Inc.](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2581)
- [Thalamuse](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2582)
- [ShowOnce](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2584)
- [Momentum](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2585)
- [Modus](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2586)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2583. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt