AI Fitness Coach MVP: Form-Check, Live in 3 Weeks
SpeedMVPs built PeakForm's AI fitness coach app MVP with form-check in three weeks: a React Native app backed by a FastAPI inference service running MediaPipe Pose on the server, streaming landmark scores over WebSockets for sub-200 ms rep feedback. GPT-4o generates adaptive weekly workout and macro plans from onboarding biometrics, goal tags and logged sessions stored in PostgreSQL with pgvector, with the stack deployed to Vercel edge functions and a GPU-backed modal.com endpoint.
SpeedMVPs built a React Native app backed by a FastAPI inference service running MediaPipe Pose on the server side, streaming landmark scores over WebSockets for sub-200 ms rep feedback. GPT-4o generated adaptive weekly workout and macro plans from onboarding biometrics, goal tags, and logged session data stored in PostgreSQL with pgvector for workout-similarity retrieval. The entire stack, inference, LLM orchestration, auth, and push notifications, was deployed to Vercel edge functions and a GPU-backed modal.com endpoint within the three-week window.
Book a free 30-minute callProject Overview
What We Built
- SpeedMVPs built a React Native app backed by a FastAPI inference service running MediaPipe Pose on the server side, streaming landmark scores over WebSockets for sub-200 ms rep feedback. GPT-4o generated adaptive weekly workout and macro plans from onboarding biometrics, goal tags, and logged session data stored in PostgreSQL with pgvector for workout-similarity retrieval. The entire stack, inference, LLM orchestration, auth, and push notifications, was deployed to Vercel edge functions and a GPU-backed modal.com endpoint within the three-week window.
The Challenge
- PeakForm's founder had a validated waitlist of 1,200 users but no working product: their in-house attempt stalled after two months on MediaPipe integration and a backend that couldn't score reps in real time. Investors wanted a functional demo in 30 days or the seed round conversation was off the table. The team needed computer-vision form analysis, an LLM-driven coaching layer, and a mobile-ready UI shipped together: not sequentially.
Our Solution
- Real-time computer-vision rep counter and form-score overlay using MediaPipe Pose streamed over WebSockets
- GPT-4o adaptive coaching engine that rewrites weekly plans based on logged fatigue, missed sessions, and progress deltas
- Macro and meal-timing planner with pgvector-powered food-swap suggestions from a 12,000-item nutrition corpus
- Biometric onboarding flow (goal, fitness level, injury flags) feeding a RAG retrieval pipeline for personalised program seeding
- Push-notification nudge system with dynamic copy generated per user's session history and streak data