ReForm

Inspiration

One of the biggest challenges beginners face when starting to exercise is with bad form. Bad form has many inverse effects such as not targeting the muscle groups effectively, or causing injury long-term. For this reason, many beginners prefer having trainers or friends who work out help them when they initially start? But what resources exist for those with access to trainers or others who work out, and still want to start their exercise journey. ReForm helps bridge this gap.

What it does

ReForm uses computer vision models (specifically MediaPipe Pose) to create a human pose estimation. It then uses these markers to determine temporal changes in your movement as your reps and set progresses. Then, based on your particular workout, it scores you based on the depth and range of your movements (this score is done deterministically based on hitting thresholds in ranges of motion). Then, any notes on your progress and set are inputted to an LLM to give human-like feedback that you can implement in your next set, as well as a graph detailing your score progression.

What technologies did you use? Did you use AI?

I used Next.js to create the web application. I used MediaPipe Pose for real-time pose estimation and tracking). I used Gemini API for the real-time text to speech, as well as for the LLM model to generate a post workout summary. Finally, I used Vercel and GitHub for deployment and version control.

I used AI in the application itself, as I used Google Gemini to generate the post workout summaries and text to speech, as well as computer vision in MediaPipe Pose for the pose tracking. I also used AI for development. I used Codex to rapidly generate code for my prototype. I utilized spec-driven development as a framework to guide the model, and used a few open source agent skills to help with my front-end development.