# Streetwise AI

- **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:** [Mustapha Abssi](https://cerebralvalley.ai/u/Mustapha)
- **GitHub:** https://github.com/abssi-maker/corner-case
- **Demo video:** https://www.youtube.com/watch?v=Y-c84ndtHCU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2420

Corner Case — Test the fix before the next crash.

City engineers have money for a few dangerous corners, not all of them. Today they judge a fix by waiting about three years for new crash data. On the stretch of Manhattan's 7th Avenue we studied, 165 pedestrians have been hit since 2023, and three were killed. Corner Case tests a fix before it is built, then checks it with a camera within weeks.

The workflow has four steps: watch the street, test fixes in a street model (a simulator of all 30 corners), try to break the plan, and set what to measure after construction.

Four Meta models each do a different job:
- SAM 3.1 tracks every pedestrian and vehicle in real Times Square footage.
- Muse Spark describes the scene and proposes what SAM should look for. SAM's tracks then check Spark's claims, so the two models verify each other.
- Muse Glimmer, Meta's open model, is fine-tuned on our street model and served on Modal. It screens all 30 intersections in 11 seconds and gives each agent a starting recommendation.
- Muse Wonder runs 30 intersection agents, each trying to beat Glimmer's recommendation in the street model, plus a coordinator. The coordinator enforces the budget and the delay limit, and rejects any fix that moves danger to the next corner.

The key moment is a plan that fails. Five stress scenarios test it: rain at night, theatre crowds, holiday crowds, a new bus lane, and drivers ignoring the new rules. Two of them, theatre crowds and rain at night, break it, and the agents revise their fixes. With equal simulation budgets, greedy search finds a larger simulated benefit, but it makes five neighbouring corners worse and fails every stress scenario. The agents' plan covers 9 corners for about $94,600. It cuts simulated harm-weighted risk by 13%, adds 3% car delay, and makes no corner worse in any of the five scenarios.

We test our own system the same way. On crash records it never saw, the street model ranked intersections poorly, below traffic volume alone. That changed the design: real injuries decide where to act first, and the street model decides which fix to pilot. The output is a decision an engineer can check: build three corners first, count close calls on camera every week, and fund the rest only if they fall within six weeks.

Glimmer trained on 3,000 scenarios generated by the street model in about 15 minutes, for under $2. On 300 scenarios from six intersections held out of training, it picks the street model's best fix 51.3% of the time, up from 21.7%. That is barely above the simple rule "always daylight the corner" (48.3%), and we say so. The real gain is elsewhere: it rates risk correctly 78% of the time (the simple rule manages 27%), returns valid output every time, and answers in 2 seconds instead of 25. Because the model is open, any city can run it on its own hardware.

## More from Meta Global AI Developer Hackathon

- [Quoc Vu Junior](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2417)
- [DigitUP](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2418)
- [understudy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2419)
- [ExplainCut](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2421)
- [Getman](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2422)
- [Clause Team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2423)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2420. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt