Building an AI Gym Copilot

Building HeliX started with a simple observation: the best personal trainers don't follow rigid programs. They adapt in real-time based on how their athlete is performing, feeling, and recovering.

From Concept to MVP

The initial prototype was built over a weekend hackathon. A simple React Native app with a camera feed that attempted to track barbell path using MediaPipe. It was rough, but the core idea clicked.

Computer Vision Pipeline

The first technical challenge was real-time pose estimation on mobile devices. We needed sub-100ms inference to provide meaningful feedback during fast, explosive movements like clean & jerks.

The Training Model

Rather than prescribing fixed rep schemes, HeliX uses an autoregulatory model based on velocity-based training principles. By tracking bar speed, we can detect fatigue before the user even feels it.

Lessons Learned

Building an AI-powered fitness app taught us that the hardest problems aren't technical — they're about understanding human behavior and building trust. Users need to believe the AI knows what it's doing before they'll hand over control of their training.

The journey continues. HeliX is currently in beta with a small group of dedicated lifters, and the results have been remarkable.