# Snap Part

- **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:** [Bibek Sarkar](https://cerebralvalley.ai/u/Bibek_01), [Bhaswati Roy](https://cerebralvalley.ai/u/Bhaswati_01)
- **GitHub:** https://github.com/bibeksarkar001/snappart
- **Demo video:** https://youtu.be/nq1fXeKzg7A
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3505

SnapPart turns three short phone videos of a physical part (a broken one, or an identical copy of a missing one) into a measured, 3D-printable STL, and says plainly what it could not see. You print an A4 mat of 26 ArUco markers and film the part: a high orbit, a low orbit, and one flipped over. The mat gives every frame its camera position and real millimetre scale. SAM 3.1 cuts the part out of every frame, a voxel carve builds the shape from the outlines, and flat faces and round bars are snapped true. Anything a camera cannot see, such as how deep a slot goes, becomes a "show me where" question, and your answer rebuilds the model in about 2 seconds with no AI call. Every value is labelled Measured, You entered, Estimated or Can't be seen. On our verified test part, a briefcase lock plate (51.6 x 41.9 x 11.8 mm by callipers), SnapPart was within 0.26 mm on every dimension and scored 0.95 on held-out views; the printed copy fit the real lock after light sanding. We also publish all 8 calliper-tested parts, failures included, in testing/. Anyone can try it with no key: the app replays our real recorded run.

Public GitHub Repository:
https://github.com/bibeksarkar001/snappart

3-Minute Demo Video: paste your YouTube link here (public or unlisted, under 3:00, tested in a private window).
The video should show, in this order: the broken part and the problem (first 15 seconds), filming three videos on the mat, the live or recorded build, the "show me where" question and the 2-second rebuild, the 3D model and honest labels, the printed part fitting, then the repo and the Meta models.

## More from Meta Global AI Developer Hackathon

- [monofilm](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3502)
- [Strativu](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3503)
- [TraceLoop](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3504)
- [FactLens AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3506)
- [Trace](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3507)
- [Ghost Agent](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3508)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3505. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt