← All projects
AI1st in state2024

Eye Don't Lie

Diabetic retinopathy screening app

Real-time diabetic retinopathy screening on an iPhone, from a live camera feed, with a robotic arm holding the phone. I built it in 8th grade.

The app running on an iPhone held by a robotic arm, next to a fundus image on a tablet
1stFlorida State Middle School Science Fair, 2024
On-deviceCore ML on every camera frame, no server
60%validation accuracy (18 of 30), so a proof of concept

Why

Diabetic retinopathy is a leading cause of blindness in working-age adults, and it's treatable when caught early. But screening takes a retinal camera and a specialist, and many diabetic patients don't get regular eye exams. I wanted to see how far a phone could get on its own.

He is demonstrating an app he created that uses the camera to diagnose an eye disease. It took first place in the state's middle school science fair.
Associated Press, May 2024

How it works

  • Capture. The flash lights the retina through a condenser lens, and the arm holds the phone at a repeatable distance.
  • Classify. Each frame goes into a Core ML model on the phone: healthy, or stage 1 to 4.
  • Guide. A dedicated class spots badly framed images and asks the user to move closer.
Capture optics, from the poster.
Capture optics, from the poster.
The phone on the Dobot arm.
The phone on the Dobot arm.

Limitations

  • A proof of concept, not a medical device. No clinical testing; tested on public fundus images, mostly photographed off a screen.
  • The app's model was validated on 30 images (5 per class). It's reliable at spotting bad framing (5 of 5), but DR-stage recall ranges from 2 of 5 to 3 of 5.
  • Not rebuilt since 2024.