# TULA

- **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:** [Suryansh pratap Singh](https://cerebralvalley.ai/u/copyninja07)
- **GitHub:** https://github.com/CopyNinja07-07/tula
- **Demo video:** https://youtu.be/q64mBXlkEQk
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/734

At an Indian FASTag toll, cars still slow down and wait. Abroad, they drive under a camera at full speed and get the bill later. India is now moving to that barrier-free system, and that raised a question for me: if no person checks the bill, who catches the mistakes? One wrong digit in a number plate sends the fine to a stranger, and in India an unpaid notice doubles after 72 hours.

TULA is a safety check that sits between the camera and the bill:
1. Meta's SAM 3.1 finds the number plate and the car in the photo.
2. Two readers read the plate separately: an open-source plate reader and Meta's Muse Glimmer 30B. Glimmer also tells me the car's make, colour and type, which is how a cloned plate on the wrong car gets caught.
3. TULA checks both reads against the FASTag and the registered vehicle and gives each car a doubt score.
4. Then it decides: charge, wait for the next camera, or send to a person with a written reason.

Today's systems charge whenever the model says it is "90% sure", but being sure is not the same as being right. TULA lets the toll operator set a promise, for example "at most 5 wrong bills in 100", and uses conformal risk control to keep it, separately for fog, clear weather and Indian plates.

I tested it on 1,050 real plate photos, including 232 foggy ones I labelled by hand. Today's approach billed the wrong person 17% of the time, and a single "90% sure" rule reached 12% in clear weather. TULA stayed under 5% in every group while still charging 86% of cars straight away. I also attacked it with cloned plates and swapped FASTags, and it stayed within its limit. The photos and both models' answers are real; the FASTag and registry data are simulated, because no public dataset links them.

Website with two live demos: https://copyninja07-07.github.io/tula/

## More from Meta Global AI Developer Hackathon

- [3D-architect](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/731)
- [SignalDesk](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/732)
- [Causewaybay AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/733)
- [Holdmate](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/735)
- [UniHub Innovators](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/736)
- [SentinelX](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/737)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/734. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt