# Gradient Descendants

- **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:** [Asish Patra](https://cerebralvalley.ai/u/ashu1123)
- **GitHub:** https://github.com/ultron513/pit-wall
- **Demo video:** https://www.youtube.com/watch?v=bhSKVI1F1a8
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/927

Pit Wall is an AI Formula 1 race strategist that answers the hardest call on the pit wall: box now, or stay out? Pick a real race, driver and lap, and an agent built on Meta's Muse Glimmer 30B (running locally through Ollama, $0 API cost) calls five tools: race overview, driver situation, a strategy simulator that searches every legal one- and two-stop plan, a fitted tyre pace model, and a rival undercut check. It returns a fixed answer (the call, pit window, confidence, rival risk), and a verifier checks that every number in it came from a tool, sending it back for a fix if not.

To test whether the strategy is any good, I built a leak-free backtest: each race is cut at the decision lap, so the strategist sees only last year's tyre data plus this year's laps so far. Across 29 first-stop decisions in six 2024 races, Pit Wall's median loss versus the best stop in hindsight was 0.23 s, against 0.39 s for the real teams, and it was within 0.5 s of the best call 21 times versus 16. In Miami 2024 it told Leclerc to pit on lap 26 (lost 0.46 s); Ferrari pitted on lap 19 (lost 3.07 s). It is not better everywhere: in Saudi Arabia last year's tyre data misled it, its mean loss (1.75 s) is higher than the teams' (1.37 s), and the hindsight judge is our own pace model. The full report, misses included, is in the repo.

Built from scratch during the hackathon: FastF1 data layer, pace model and simulator, 5-tool agent with verifier, backtest, FastAPI web app with live streaming and replays, and 55 tests.

## More from Meta Global AI Developer Hackathon

- [DUKERS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/924)
- [UFAT](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/925)
- [atlas](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/926)
- [Lucas Futures](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/928)
- [Jones Melton](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/929)
- [Burotu](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/930)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/927. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt