Train Anywhere

Multi-device AI-powered fitness companion with real-time workout logs, form analysis, contextual coaching, and training data syncing. Track every rep in real time without extra bulk in your gym bag.

Comment

Inspiration

The fitness industry has long struggled with a fundamental problem: traditional workout and fitness tracking is an exercise of its own. Most fitness apps are about as useful as an empty notebook, relying on manual logging, which is time-consuming and provides no insight into actual performance and execution. Meanwhile, personal trainers—who can provide real-time feedback and corrections—are prohibitively expensive for most people. We wanted to democratize access to professional-grade exercise and fitness analysis, and have been testing this concept as our "Will Smith Eating Spaghetti" to asses performance improvements to coding and app building agents in 2025.

What it does

Train Anywhere is an AI-powered fitness companion that combines advanced computer vision with Gemini 3's multimodal reasoning to provide professional-grade workout tracking and coaching for 11k+ movements, body composition scanning, and other health and recovery metrics through your phone and laptop (no extra hardware required).

Core Features:

  • Real-Time Computer Vision Analysis: Tracks 33 skeletal points across 11k movements to analyze rep tempo (eccentric/concentric phases), depth (e.g., hip crease below knee for squats), and range of motion
  • Contextual AI Coach: Powered by Gemini 3, remembers every rep you've done and provides personalized, context-aware coaching based on your historical performance
  • Unified Data Hub: Upload and centralize all training data with plans to import from handwritten notes and other apps
  • Community Features: Global challenges and leaderboards to keep users motivated

How we built it

Coding Agents: Google AI Studio + Antigravity Frontend: Next.js 14 with React, TypeScript, and Tailwind CSS for a responsive, modern UI Computer Vision Pipeline: MediaPipe Pose for real-time skeletal tracking at 30fps, analyzing movement patterns with sub-degree precision AI Integration: Gemini 3 Pro powers the contextual coaching system through structured prompts that include workout history, current performance metrics, and user goals Backend: Node.js API with PostgreSQL for data persistence Real-time Processing: Optimized video processing pipeline to run on-device, ensuring privacy and zero latency

Gemini 3 Integration Highlights:

  • Used/Using Google AI studio and Antigravity
  • Processes workout context (historical data + current session metrics)
  • Generates personalized coaching advice considering user progress over time
  • Provides natural language explanations of form corrections
  • Adapts recommendations based on detected patterns in training data

Challenges we ran into

  1. Performance Optimization: Running real-time pose estimation on mobile devices while maintaining 30fps required extensive optimization of our MediaPipe pipeline
  2. Contextual Memory Management: Designing prompts for Gemini 3 that effectively utilized historical workout data without overwhelming the context window
  3. Form Analysis Accuracy: Calibrating the computer vision algorithms to accurately detect depth and range of motion across different body types and camera angles
  4. Progressive Web App Constraints: Accessing camera permissions and handling video streams in PWA format across different mobile browsers

Accomplishments that we're proud of

  • Achieved <100ms latency for form feedback on mid-range mobile devices
  • Created an intuitive UX that makes professional-grade analysis accessible to complete beginners
  • Successfully integrated Gemini 3's multimodal capabilities to bridge quantitative metrics with qualitative voice coaching
  • Scraping and enriching 11k movements to offer camera tracking and AI trainer feedback

What we learned

  • The power of combining traditional computer vision with LLM reasoning—CV provides precise measurements, while Gemini 3 contextualizes them
  • Effective prompt engineering for fitness coaching requires structured data formats and clear constraint definitions
  • Real-time ML on mobile is viable with the right optimization strategies

What's next for Train Anywhere

  • Program Generation: Use Gemini 3 to automatically generate personalized training programs based on goals, equipment, and historical performance
  • Social Features: Expand community features with AI-moderated challenges and peer coaching
  • Wearable Integration: Sync with fitness trackers for more holistic insights

Built With

Share this project:

Updates