# Mirai Climbing

- **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:** [Yuzuki Kondo](https://cerebralvalley.ai/u/tanuki_ponzu), [Ajay Ankam](https://cerebralvalley.ai/u/ajay_ankam), [si ho](https://cerebralvalley.ai/u/siho9o9o), [Paolo Valdez](https://cerebralvalley.ai/u/_pablo_vee)
- **GitHub:** https://github.com/Mirai-Climbing/topo-agent
- **Demo video:** https://youtu.be/3dhX3NPePSI
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1807

Climbing apps let climbers find a gym's problems, log their sends and share them. But first someone has to add each problem: take a photo of the wall and mark every one of its holds by hand. That takes time, and walls are reset again and again, so keeping up is hard.

topo-agent makes adding a problem one tap, so any climber at any gym can do it. Take a photo of the wall and tap one hold. SAM 3.1 outlines every hold on the wall, and a Muse Spark agent picks only the holds that belong to the same problem. The app shows what the agent did, the climber can add or remove holds, and the result is saved as a shared problem in Mirai Climbing, our climbing app, so nobody has to mark the same problem twice.

Because the problem is outlined, not just colored, it also helps climbers who can't tell hold colors apart: red-green color vision deficiency affects up to one man in twelve (men of Northern European descent; about one in twenty in Japan).

Results so far: on drawn test walls, 99% recall and 100% precision (tape problems: 14% with a color filter, 100% with the agent). On 68 routes in real gym photos, 60% recall and 68% precision; the agent still stops too early on dense walls, and every experiment is logged in the repo. We also trained a small YOLO model that replaces SAM: the same recall in about half the time, less precise for now, and the YOLO model also runs offline on Android (ExecuTorch); choosing the problem still uses Muse Spark.

Mirai Climbing existed before the hackathon. The topo agent, its evaluation and its integration into the app were built during the hackathon; pre-hackathon code is kept only as a baseline.

## More from Meta Global AI Developer Hackathon

- [PerceptaHeal](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1804)
- [Ikigai Loop](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1805)
- [Milo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1806)
- [Ai study assistant](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1808)
- [Buildlers](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1809)
- [Harrison Hlongwane](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1810)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1807. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt