# KineSight

- **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:** [Ameer Hamza](https://cerebralvalley.ai/u/_irrepressible)
- **GitHub:** https://github.com/irrepressible007/kinesight
- **Demo video:** https://youtu.be/hC6k1zX42_Y?si=eFZgUtzEZveDCU41
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4263

https://kinesight.onrender.com/

KineSight turns any ordinary overhead or bench camera into a sensorless QA station for hardware and garment assembly lines. Catching a skipped screw, a rushed solder joint or an out-of-order step normally means $10k–$25k per bench in smart torque drivers, RFID gates and PLC sensors, which factories in Bangladesh and other emerging markets can't afford. Sending raw video to a vision-language model doesn't solve it either: VLMs sample sparse frames and guess, so they can't tell a 0.4 s iron contact from the required 2.0 s, or count four screw-downs across 400 frames.

KineSight splits the job between models and math. Muse Spark 1.3 compiles a plain-text or PDF SOP into machine-checkable rules: which tool must touch which part, for how long, and how many times. SAM 3.1 segments and tracks those tools and parts through the video from text prompts, with no training. A deterministic OpenCV/NumPy engine then computes centroids, tool speed and dilated-mask contact frame by frame and turns them into timed contact intervals. When a rule fails, only three keyframes around that moment go back to Muse Spark, which confirms whether the defect is real or an occlusion and explains the root cause. Muse Image then draws a corrective work-instruction diagram.

In the demo, an operator drives only 3 of the 4 corner screws on a mounting plate. KineSight measures three contacts (6.4 s, 11.5 s, 22.8 s), flags the shortfall at t = 26.07 s (FAIL, 75% compliance), and Muse Spark confirms from the keyframes that the top-right hole is empty. The compliant run of the same station passes 4/4, and a soldering bench with a different camera angle and different objects passes both of its steps (flux dip 3.4 s, joint dwell 14.9 s).

## More from Meta Global AI Developer Hackathon

- [Bakery Intelligence](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4260)
- [speed up](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4261)
- [TAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4262)
- [FollowCut](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4264)
- [Gameon Finance](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4265)
- [RobotRanger](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4266)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4263. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt