AI Soccer Coach

# Inspiration One-on-one soccer coaching costs $80+/hr and requires scheduling that doesn't fit a high schooler's life. We wanted to democratize access — give any player the kind of specific, technical feedback that only well-funded programs can afford, available anytime, on demand. Also the world cup recently ending was a major source of inspiration.

## What it does AI Soccer Coach lets you record a drill, analyzes your body mechanics in-browser using pose estimation, and returns structured coaching feedback — what's wrong, how to fix it, and what to practice next. An LLM coaching brain also generates personalized drill plans, answers technique questions, and adapts to different playing philosophies. Progress is tracked across every session.

## How we built it React + Vite frontend with MediaPipe PoseLandmarker running entirely in-browser — video never leaves your device. A FastAPI backend computes biomechanical descriptors (plant foot angle, knee flexion, hip rotation, follow-through) from the landmark data and sends them to the Claude API, which returns structured coaching feedback. Supabase handles auth, session storage, and the progress timeline. Deployed on Vercel and Render.

## Challenges we ran into Detecting the exact contact frame without ball tracking was the hardest technical problem — we solved it using the kicking foot's velocity peak from landmark data, with a manual scrub fallback so CV issues can never block the user. Getting the LLM to return actionable feedback (not just identify the flaw, but prescribe the fix and a follow-up drill) required significant prompt engineering.

## Accomplishments that we're proud of All pose analysis runs on-device — no video upload, no server CV, no GPU. The feedback pipeline goes from raw video to specific coaching cues in under 60 seconds. We built a drill library covering 8 positions and 4 playing styles with full instructions that the LLM coaching brain can assign dynamically.

## What we learned The capture setup is as important as the model — enforcing a correct camera angle before recording was the single biggest improvement to analysis quality. We also learned that leading with the LLM coaching brain (not the CV) was the right call: the conversational coach delivers more consistent value and carries less technical risk.

## What's next for AI Soccer Coach Ball detection for shot outcome and dribbling speed (Phase 3), additional position-specific drills, health and recovery integration via WHOOP for recovery-aware training, and eventually a sport-agnostic movement coaching engine — the same pipeline applied to tennis, basketball, and physiotherapy rehab.

## Built with React, Vite, TypeScript, Tailwind, MediaPipe Tasks (Vision), FastAPI, Python, NumPy, SciPy, Claude API (Anthropic), Supabase, Vercel, Render

Built With

• claude
• fastapi
• mediapipe
• react
• render
• supabase
• typescript
• vercel
• vite