# LineSight

- **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:** [Krystian Hettman](https://cerebralvalley.ai/u/hetmix)
- **GitHub:** https://github.com/hetKr/LineSight
- **Demo video:** https://youtu.be/1zCmX6BVlW0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3975

LineSight turns a conversation, factory cameras and machine data into a calibrated production-line simulation.

The problem: most factories know a part only at its scan points, and everything in between is invisible. Classic simulation could answer "what should we fix?", but it takes weeks of modelling, so most lines run on estimates. People decide on numbers nobody has measured, and money goes to the wrong machine.

How it works:
1. Describe the line in plain words. An agent on Muse Spark Open asks what it needs and builds a runnable model. It adds only what you described, and every number records where it came from: you, an assumption, a camera or data.
2. Give it recordings from existing cameras. SAM 3.1 and Muse Spark Open propose zones at setup, and you confirm them. Computer vision then measures on your own machine: arc light, stack lights, motion and line flow, with people, signs and the date masked.
3. Add machine or MES data if you have it, as any CSV or TSV. LineSight proposes what each column means, looks for relationships, checks them on held-out data, and puts what holds into the simulation by itself. The more you give it, the closer it gets.
4. It compares the model with the measurements ("you said 30 s, the camera measured 85 s"), and you decide what to replace.
5. Ask "what if this robot were twice as fast?" or "what's the cheapest +15 %?". The agent runs simulations in parallel and answers with confidence intervals, so you know before you spend.

In the video's demo, the line's limit turns out to be the unloading robot, not the welding cells: faster welding gives nothing, and a robot 15 % faster gives almost 18 % more. A fictional press cell shows the same path on other data: the program notices by itself that rejects are ten times higher for two minutes after a fault.

Limits, stated plainly: the video's camera footage comes from a real station, with people, logo and date masked, while all the numbers and conclusions in that demo are generated to protect the plant. That example and its footage are not in the repository. The repository runs two generated examples offline, with no keys. Each generalisation sits behind a flag and a gate written before it ran, and failed gates stay recorded in ROADMAP.md. Motion-based measurement and two-camera fusion have not passed all their gates yet.

## More from Meta Global AI Developer Hackathon

- [Second Ear](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3972)
- [Vibe Doc](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3973)
- [Ginjikoibito](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3974)
- [DocSyde](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3976)
- [MonkeeTools](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3977)
- [Dmitry (Dima) Zabello](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3978)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3975. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt