BiomechAI

Inspiration

Physical therapy and fitness coaching are expensive and inaccessible to many. Most people exercise with poor form, leading to injuries that could be prevented with real-time feedback. We wanted to democratize biomechanical analysis by using only a webcam — no wearables, no sensors, no cost barrier.

What it does

BiomechAI analyzes exercise form in real time using MediaPipe Pose running entirely in the browser. It tracks joint angles (knees, hips, back, elbows), counts reps automatically, and alerts users when form deviates from safe ranges. After each session, an AI coach (powered by Nvidia NIM) generates a personalized report with strengths, improvement tips, and injury risk assessment. Users can track progress over time through a dashboard with visual charts.

How we built it

Frontend: React 18 + Vite with MediaPipe Pose for in-browser landmark detection, Recharts for progress visualization Backend: Express.js REST API with MySQL for session storage AI: Nvidia NIM API (minimax-m2.7 model) generating structured JSON coaching reports Auth: JWT-based authentication with per-user session history Analysis: Custom angle calculation engine using arctangent formulas on 33 body landmarks, with phase-detection rep counting

Challenges we ran into

Getting accurate rep counting required carefully tuned angle thresholds for each exercise — a squat "down" phase differs from a deadlift "down" phase. Balancing real-time performance while running pose detection on every frame was tricky. We also had to design a form scoring system that penalizes bad form without being overly punishing to beginners.

Accomplishments that we're proud of

Fully in-browser pose detection — no video leaves the user's device Clean phase-detection algorithm that reliably counts reps across three exercise types AI coaching reports that give actionable, personalized feedback Polished UX with onboarding modal, exercise setup guides, and a demo account for judges

What we learned

MediaPipe Pose is incredibly capable for real-time browser-based analysis. Structuring AI output as JSON (not free text) makes post-processing and display much easier. Rep counting with joint angles is as much about timing thresholds as it is about angle math. Good onboarding is critical — users need to know how to position their camera before analysis starts.

What's next for BiomechAI

Support for more exercises (lunges, planks, overhead press) Live AI voice coaching during sessions, not just post-session reports Social features — share sessions and compete with friends Mobile app with the same MediaPipe pipeline Integration with wearable data (Apple Watch, etc.) to combine visual + sensor analysis

Built With

• node.js