# Retake Coach

- **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:** [rami roobi](https://cerebralvalley.ai/u/ramayan)
- **GitHub:** https://github.com/roninram/retake-coach
- **Demo video:** https://youtu.be/qxymNLWxh-U
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1013

Retake Coach checks a fundus photo the moment it is taken and tells the technician to keep it or retake it. For a retake, it says what to fix (focus, aim, or something blocking the view) and marks where on the photo the problem is. No photo leaves the computer next to the camera. Each visit adds a row of counts to a training report, with no patient details. In diabetic eye screening, ungradable photos are common: 12.4% of eyes in one US telehealth program, and 23.9% of images with a handheld camera in India. Each one usually means calling the patient back. A built-in guide explains every word, number and mark on the page. Retake Coach checks photo quality only and gives no medical advice.

Results: On 400 held-out DeepDRiD photos, it detects focus, aim and obstruction problems with AUROC 0.90, 0.85 and 0.95. It catches 74% of the photos graders rejected and sends back 23% of good ones. On IDRiD's held-out photos, it places the optic disc and fovea a median 0.08 disc diameters from the experts' marks. With DINOv3, the map's strongest spot lands on a pasted shadow 85% and a pasted glare 80% of the time (chance: 7%). Fitted with no quality labels, the guard's energy already ranks not-good-enough photos higher (AUROC 0.72; 0.80 for obstruction).
Built during the event: Everything, from Oct 26 to Nov 4. The repository also holds a model card and 27 automated tests that run on every push. It keeps the directions we tried first too, including what didn't work.
What's next: Add a new camera from its own unlabeled photos (built for the Phelcom Eyer2 handheld camera, being tested now), then test with a screening clinic's technicians, using the visit report to measure retakes per visit.

## More from Meta Global AI Developer Hackathon

- [Study Hall Media](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1010)
- [KryomLabs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1011)
- [Paycheck Committee](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1012)
- [Openwist](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1014)
- [Team AgentArena](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1015)
- [ElasticShelf](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1016)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1013. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt