# Signoff

- **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:** [Kranthi Sai Gadi](https://cerebralvalley.ai/u/Kranthi_007)
- **GitHub:** https://github.com/Kranthisai07/SignOff
- **Demo video:** https://youtu.be/HepkUTl71zw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2748

Signoff gives anyone the confidence to do things they've never done before, with Meta watching their back.

Ask your Ray-Ban Meta glasses for help with any hands-on task: clearing the ice off your driveway, organizing a shoe rack, fixing a wobbly bed frame, assembling a desk. Signoff's idea is that Meta reads what the task needs (the manual, the instructions online), breaks it into steps, and writes its own sign-off list: for every step, a concrete picture of what "done right" looks like. Then it walks you through the task from the first step to the last. Because the glasses see what you see, Signoff doesn't just tell you the next step. It checks the step you just did, and signs it off only when it's actually right. A wrong step is caught, with a reason, before it turns into a broken desk or a wasted afternoon.

That's the difference: assistants talk you through a task. Signoff makes sure you got it done.

What we built in this hackathon is the heart of that loop, working end to end on real hardware:
- Step checks: every step carries an instruction and a "done right" criterion.
- Glasses capture: an Android app on Meta's Wearables Device Access Toolkit checks a step from a photo taken on Ray-Ban Meta Gen 2.
- Verification: Muse Spark judges the photo against that step's check and answers pass, fail or unsure, always with a reason, in about 4 seconds.
- A benchmark, so the claims are measured, not hoped.

We proved it on a first experiment, a five-step fried egg, chosen because right and wrong are quick to stage: 29 of 30 photos judged correctly and 0 of 14 wrong steps signed off (20 photos first-person from the glasses, 10 of them live through the app).

Next: the guided start ("Hey Meta, help me assemble this desk"), sign-off lists generated automatically from manuals and web instructions, and checking steps from continuous video. The fried egg was the experiment. The goal is any task with steps you can see.

## More from Meta Global AI Developer Hackathon

- [HackPulse](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2745)
- [MultiplAIers](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2746)
- [Oral-lens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2747)
- [RoadReport](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2749)
- [BLE Jarvis](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2750)
- [VibeSolo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2751)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2748. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt