# Saad Mansour · SOLANG

- **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:** [Saad Mansour](https://cerebralvalley.ai/u/saadmansour)
- **GitHub:** https://github.com/saad-mansour/solang-vision-qc
- **Demo video:** https://youtu.be/8AhQ_bW44q0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/363

SOLANG Vision QC — a second pair of eyes on the final-inspection table of a children's-wear factory in Cairo.

The problem, firsthand: every garment passes a human final inspection before packing. After a few hours the eye gets tired, small defects slip through, and the order comes back from the customer with shipping paid both ways. Our own estimate is that roughly 15% of orders come back with a defect involved (no audited figure — our shop platform does not report returns); the goal is to bring that under 5%.

What it is: an ordinary phone on a stand above the inspection table and a tablet in front of the quality inspector — one link, nothing to install, in the inspector's own language (Arabic; English is one tap away). One tap on the tablet: the phone shoots at full resolution, SAM 3.1 finds the garment, Muse Spark inspects it tile by tile, re-checks every candidate magnified with the inspector's past verdicts as examples, looks at prints and cuffs up close, compares the two sleeves and the overall look against a reference piece. The inspector sees circles on the photo with a reason, confirms or rejects each one, marks what the camera missed, names who sewed it and decides pass / repair / reject. Her verdicts are stored as labelled crops and ride along in the next inspection of the same product — the system suggests, the inspector decides, and it learns from her.

Built from zero during the hackathon (Oct 26 – Nov 4). Pre-existing: only the factory and its paper QC sheet. Used on the real table from day four: 38 inspections, 70 spots shown to the inspector, 38 confirmed and 32 rejected (precision 54%), 10 misses she marked (recall 79%); loose threads are the strongest type (84%), stains the weakest (32%); about 70 s per piece on the factory Wi-Fi and about $0.18 per piece. Every number, every miss and the sleeve-length check next to tape measurements are in the README.

Live demo (no key needed, three bundled factory photos to try): https://solang-vision-qc.saadmansour.workers.dev/station/

## More from Meta Global AI Developer Hackathon

- [BlinkFlow](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/360)
- [PulseGuard](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/361)
- [ARIQ Water](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/362)
- [thrqsytr](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/364)
- [WrenchVision](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/365)
- [Oddly Clear Studio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/366)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/363. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt