All work Case study · Thesis

RevAItalize

A post-rehabilitation coach for lower back pain. The webcam tracks your body, an AI model checks your form, and you get feedback on every rep.

My role
Front-end, UI/UX, integration testing
Timeline
2024 - 2026
Team
3 developers
Type
Computer Science thesis
revaitalize
RevAItalize landing page: A post-rehabilitation for your lower back pain

The problem

After physical therapy, most patients keep exercising at home, alone. Without a therapist watching, it's easy to do a stretch wrong for weeks, which slows recovery and can cause re-injury.

RevAItalize gives patients a coach that's always there. It needs only a laptop or phone camera: no appointment, no wearables, no special equipment.

How it works

  1. Set upThe app checks that your whole body is in frame before the session starts.
  2. Follow the guideA guide video plays beside your camera, so you can match the movement.
  3. Get checkedBlazePose finds your body keypoints on every frame. An LSTM model reads the last 20 frames and flags the form errors it sees.
  4. See progressScores land on your dashboard as trends and history, with a downloadable PDF report.

Try a session

A simplified version of the session screen. Press Start to count reps; some reps come back with a form warning, just like the real thing.

revaitalize · session
RevAItalize
Dashboard
Profile
Session
Torso Rotation
Your camera
Guide video
0
Set 1 of 2Reps
Press Start to begin
Session accuracy: 0%

My part

I owned the front end and the UI/UX: designing the screens, building them in React and TypeScript, and testing the app against the back end and model as they came together.

Design

Landing page, onboarding, dashboard and session screens, built for patients who aren't tech-savvy.

Build

Components on shadcn/ui and Tailwind, charts with Recharts, and motion for transitions.

Session screen

Camera and guide video side by side, a rep counter, live feedback and an accuracy bar.

Integration testing

Checked the full loop with the team, from camera frames to model results to saved sessions.

Exercises

Three seated exercises for lower back recovery, each with its own guide video and its own tuned form checks.

01Flank Stretch
02Hiding Face
03Torso Rotation

Under the hood

  • BlazePose. Tracks body keypoints from the webcam on every frame, right in the browser.
  • LSTM model. A long short-term memory network, trained in TensorFlow, learns how each exercise moves over time. It reads a window of 20 frames and flags up to six form errors, with thresholds tuned per exercise.
  • Our own dataset. We recorded the training data ourselves through a recording tool that streamed keypoints to the server.
  • API and data. FastAPI serves the model, PostgreSQL stores users and sessions, and Docker Compose runs it all together.
  • Hosting. Front end on Vercel, with the back end reached through a Cloudflare Tunnel.
React 19TypeScriptViteTailwindshadcn/uiRechartsBlazePoseLSTMTensorFlowFastAPIPostgreSQLDocker

Screens

Outcome

How the thesis defense went, and what you learned working on a team of three.

Next projectScrapNote