# Crosscheck 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:** [Shubham Srivastava](https://cerebralvalley.ai/u/shubham2050)
- **GitHub:** https://github.com/bravoshubham/crosscheck
- **Demo video:** https://www.youtube.com/watch?v=gayHcEkgkPc
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3328

Air traffic control clears a 737 to 11,000 feet. The crew reads it back perfectly, then dials 12,000 into the autopilot. Nothing on the radio sounds wrong, and nobody notices. Crosscheck does. It's watching the panel through a camera, and it says "Cleared one-one thousand, set one-two thousand" 12 seconds before the plane would have flown through 11,000.

On an airliner, one pilot flies and the other watches, checks and speaks up. That second job is the one people do worst when they're tired or busy. When the NTSB looked at 37 airline accidents where the crew played a part, someone failed to monitor or challenge in 31 of them. That's the gap I built Crosscheck to close.

Crosscheck is a colleague that only watches and listens. It hears the radio and the intercom, reads the flight displays through a camera, and never plugs into the aircraft. When what was said, what was set and what the plane is actually doing don't match, it says so in the same few standard words a good first officer would use. Plain code decides when to speak. Meta's models do the hearing, the seeing and the explaining. It never touches the controls. The crew decides.

I flew every test live in a Boeing 737 in X-Plane 12. It catches:
- an altitude set differently from the clearance, or read back wrong
- an altimeter setting different from the one ATC gave
- the autothrottle dropping out on final approach, the pattern behind Asiana 214, before the speed is 10 knots low
- flaps set differently from what was called
- a captain who slumps and doesn't answer: it asks first, then tells the first officer to take control (an opt-in camera, tested on generated footage)
- and it stays quiet when nothing is wrong: fifteen minutes of normal flying, every line spoken and transcribed live, no false alarm

None of these checks is specific to the 737. They're the cross-checks every airline crew makes, and because Crosscheck reads the displays the way a pilot does, moving to an A320, a 787 or an old analog cockpit means teaching it those displays, not wiring anything in. I think it could matter even more for people who fly alone, like private pilots and hobbyists with no second pilot at all, and for flight schools, to check students' callouts and debrief every lesson. That's where I want to take it next.

Every run is scored automatically against the simulator's own data, which Crosscheck never sees. 104 of 112 live runs came out right, and 29 of 29 on board with the network cut off. The report shows every run, misses included, with the camera frame from the moment Crosscheck spoke. All the recordings are in the repo, and one command replays all 27 without a simulator, GPU or API key.

Live report: https://bravoshubham.github.io/crosscheck/

## More from Meta Global AI Developer Hackathon

- [Clever Monkey](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3325)
- [Neuron](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3326)
- [Apprentice - Taran Pal Singh](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3327)
- [Speed&Inertia](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3329)
- [MYB](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3330)
- [Reality Fork](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3331)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3328. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt