← Back to work

02 · Software Engineering Intern · 2023

TG3D Studio

ML-driven fitness — form scoring & personalized training

Overview

Applied machine learning to turn exercise-app movement data into useful, personalized coaching.

As a summer intern I owned two ML-driven features end to end — from model training and evaluation through the backend and the frontend customers touched — using Ruby on Rails, Docker, and fine-tuned language models. The goal was practical: help people work out better by telling them whether they're doing an exercise correctly and giving them a plan built around their body and goals.

Below are the two pieces I shipped and how I validated them.

01

Form Scoring

Trained models on real movement data to score how correctly someone is performing an exercise.

Workout effectiveness comes down to form — the same rep done wrong is wasted or unsafe. Using movement data captured from people using the exercise app, I built a model that classifies whether an exercise is being performed correctly, turning raw motion into a usable effectiveness signal the product could act on.

  • Collected and prepared movement data from real app users as training input.
  • Trained and iterated on the classifier, evaluating accuracy against held-out examples.
  • Integrated the form signal into the app so it could give users live feedback.
50→95%correct-form identification accuracy
02

Personal Plans

Built a feature that generates a training regimen tailored to each user's body and goal.

Beyond scoring individual reps, I shipped a personalized regimen generator end to end. I fine-tuned GPT and Llama 2 to produce training plans from each user's height, weight, and target outcome, and used few-shot prompting with well-researched reference exercises to keep the output relevant and safe. The whole feature — frontend and backend — went out as a brand-new capability in the app.

  • Fine-tuned GPT and Llama 2 to generate regimens from user height, weight, and goal.
  • Applied few-shot prompting with researched exercises to raise output relevance.
  • Implemented the frontend and backend with Ruby on Rails and Docker.
+70%improvement in output relevance
>95%correctly formatted responses across 200+ runs

Stack