# metalmind

- **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:** [Rami Kamel](https://cerebralvalley.ai/u/ramik)
- **GitHub:** https://github.com/ramikamel/metalmind
- **Demo video:** https://youtu.be/JbjNXGTT2yA
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3712

MetalMind is an AI agent that helps PyTorch developers get more performance out of Apple Silicon without writing Metal kernels by hand. It takes a supported PyTorch function and tensor signature, asks Meta’s Muse Wonder to propose a Metal kernel, then compiles and runs the candidate on the Mac. Compiler errors, correctness failures, and benchmark results inform the next proposal. MetalMind saves a kernel only after it passes held-out correctness tests and beats the fastest verified PyTorch baseline in two independent timing sessions. Saved kernels run offline, and incompatible calls fall back to PyTorch.

In the demo, MetalMind rejects a fast FP16 softmax kernel because identical inputs produce different outputs. With targeted synchronization feedback, Muse proposes a replacement that passes 22 held-out cases and runs 3.10× and 3.18× as fast as eager PyTorch in two sessions on a 14-core M3 Pro, with caller-guaranteed input bounds. The demo then runs the saved kernel on fresh inputs with zero model requests and shows PyTorch fallback when the shape, dtype, or gradient requirements change.

## More from Meta Global AI Developer Hackathon

- [Tokenmaxxing](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3709)
- [ChronoRig](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3710)
- [Wazi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3711)
- [Vantiq Studio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3713)
- [WorkLens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3714)
- [Machine Vision Radar](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3715)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3712. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt