AI-powered home physical therapy — real-time form feedback, no clinic required.
🏆 Winner, 2025 Congressional App Challenge (NV-04, Rep. Steven Horsford) · Announcement 📺 Demo video
A teammate's grandmother broke her wrist and was prescribed physical therapy twice a week for six weeks. Long commutes and limited mobility meant she missed most of her appointments — and when she tried the exercises at home, alone, she did them wrong.
1 in 2 adults needs physical therapy at some point. Most insurance plans cover only 20–60 sessions a year, leaving the elderly and uninsured to recover on their own, with no feedback.
Flexeon gives them that feedback.
Point your phone at yourself and start an exercise. Flexeon tracks your body in real time and tells you whether you're doing it right.
- Live pose tracking — full-body skeleton overlay rendered on the camera feed
- Form correction — joint angles are compared against target ranges for each exercise, with on-screen cues when you're out of range
- Rep counting — automatic, based on movement through the target range
- Fully on-device — video never leaves the phone
Camera frames are captured with react-native-vision-camera and downscaled in a frame processor (vision-camera-resize-plugin), then run through a MoveNet pose-estimation model via TensorFlow Lite (react-native-fast-tflite).
Inference runs inside a worklet on a separate thread, so the camera preview stays smooth while the model executes on every frame. The resulting keypoints are converted into joint angles, which drive the form feedback, rep counting, and skeleton overlay.
Stack: React Native (Expo SDK 54) · TypeScript · VisionCamera · TensorFlow Lite / MoveNet · Worklets · Firebase Auth · Socket.IO
Flexeon needs a development build — the native ML modules (VisionCamera, fast-tflite) don't run in Expo Go.
npm install
npx expo run:android # or: npx expo run:iosRequires a physical device with a camera. Grant camera permission on first launch.
Built with Elijah Ladot and Xuanzhe Wang.
My role: I built most of the app — the VisionCamera → TFLite inference pipeline, the pose-angle feedback and rep-counting logic, and the UI.