# Parallax

- **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:** [Maruthi Kunchala](https://cerebralvalley.ai/u/King_Maku)
- **GitHub:** https://github.com/KMaruthi2002/EIDOS
- **Demo video:** https://youtu.be/_JSD6NeWwkg
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1391

EIDOS turns one ordinary video of someone doing a task into a skill that a camera, or a robot, can check.

People film themselves cooking, wiring, assembling and repairing machines every day. There are billions of hours of it, but AI still can’t watch one of those videos and come away with a skill it can check. Today’s AI assistants can see what you see and answer when you ask. EIDOS knows what should happen next, and speaks up the moment something goes wrong, before you ask.

How it works:

Learn from one video, with no labels. Muse Spark watches the video eight seconds at a time. Muse Wonder compiles what it saw into a program: the steps, the order they really need, amounts you can change, and a visual check for every step. SAM 3.1 outlines what each check is about.
Follow along live. Point any camera at the work. EIDOS knows which step you’re on, says what comes next, and checks each step against the video. It raises an alert the moment a step is skipped or done out of order, and it looks back over the recording before accusing anyone. Example: learned from one technician’s video, it flags “part removed before the power was switched off” while it is happening.
Any task, any machine. The same skill format covers cooking, PC hardware, electrical maintenance and a robot arm. In simulation, the arm is shown a task once, checks every step through its cameras, and rebuilds its work when it gets knocked over. With checks off, it says it’s done next to a broken stack.
Works without a key. Practice recordings come with recorded runs of Muse Spark, so anyone can watch EIDOS follow a task live in the Studio with no API key.

Tested on CaptainCook4D, real people cooking with their mistakes labelled. Running live on recipes it had never seen, EIDOS named skipped steps as they happened, and it falsely accused correct cooks about four times less often than Muse Spark looking over the whole video. The README has the full results, including what is not solved yet.

Everything ships as EIDOS Studio, an open-source web app.

## More from Meta Global AI Developer Hackathon

- [DoodleByte](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1388)
- [Inference Exchange](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1389)
- [shaik](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1390)
- [By My Side Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1392)
- [Xiling](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1393)
- [Slopbook Relay](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1394)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1391. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt