# NearMissLens

- **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:** [Shahraz Hussain Siyam](https://cerebralvalley.ai/u/shahraz_23)
- **GitHub:** https://github.com/Morningstaaar36/nearmisslens
- **Demo video:** https://youtu.be/WjMjqWtJCO4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3053

NearMissLens checks a proposed street safety fix against the city's own crash record before the money is spent. Give it a project, like a leading pedestrian interval at one intersection, and in about a minute it tells you whether the crashes at that corner are the kind this fix prevents, whether the fix is already installed, whether the city's own past program of that fix worked, and whether another fix would prevent more. A review costs about 8 cents.

US cities are spending billions in federal safety money, yet the lists of fixes are often drawn up before anyone has looked at the corner itself. NearMissLens caught exactly this in real proposals. A Washington DC neighborhood commission asked for a leading pedestrian interval at 22nd Street and L Street NW, but DC's own signal records show one is already installed. NYC DOT proposed intervals at three Lafayette Street corners, and New York's signal dataset shows one has been running at Prince Street since October 2024.

The tool does four jobs. It reviews a single project and returns a verdict with its reasons, plus a pre-mortem written by a second model that argues against the verdict. It reviews a whole program: Muse Spark searches the web for the agency's own project pages, reads the projects off them, and code ranks every one within the user's budget. It checks a claim from a news story or an official against the crash record, number by number. It also builds a full report for any intersection, ending in a drafted request letter to the agency that owns the road.

What sets it apart is the division of work. Meta's models do the searching, reading and writing, while code sets the verdict and checks everything before a user sees it. Every number must match the record, and every quote must be found word for word on the page it cites. Whether a city program worked is decided by a causal test with placebo dates that neither the model nor a user can override.

Every test was registered in the repository before it ran, and we report the misses alongside the wins. On 36 fresh real cases from the live web, 33 passed the gate and all 72 of their quotes were found on the cited page, at 6 to 13 cents a case. Given the same inputs and the same web search but none of our tools, Muse Spark cited 26 of 73 quotes that were not on those pages. The tool reviewed 7 of 7 real proposals from 5 cities on the correct corner and 21 of 23 projects across 3 real agency programs, and it found all 16 hand-collected projects on agency web pages. Its causal test measured a 17 percent drop in pedestrian injury crashes from New York's leading pedestrian intervals; a naive before and after comparison claims 24.

It covers New York, Washington DC, Seattle, San Francisco and Chicago, and adding a city is a data entry, not new code. It is live and free to try, with no account or key.

Live: https://hussainshahraz720--nearmisslens-copilot-web.us-east.modal.run
Results: https://morningstaaar36.github.io/nearmisslens/

## More from Meta Global AI Developer Hackathon

- [EdgeMind.AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3050)
- [PairProof](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3051)
- [Christian Friday Douglas](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3052)
- [AU404](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3054)
- [remy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3055)
- [PreFlight_PCB](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3056)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3053. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt