GymBro AI

Main dashboard user sees with user name at top
Selection screen for type of workout
Graph displaying workout stats for this session and the week
Example of end position for workout
Example of start position for workout

Inspiration

We are all freshmen at Stevens Institute of Technology. Earlier this semester (this Thursday), part of the roof of our campus fitness center collapsed, which forced the gym to shut down entirely. On top of that, working out alone can already feel intimidating, especially for students who are new to college.

We wanted to build something that makes working out feel less lonely and more motivating. Our goal was to recreate the feeling of having a gym partner with you, even when you are training by yourself.

What it does

GymBro AI is an AI-powered workout buddy that uses computer vision to track exercises and count reps in real time. It provides live voice feedback through a personality-driven AI coach.

Users can choose between:

Nice mode, which is supportive and encouraging
Mean mode, which playfully teases the user in a safe and motivating way

After each workout, the app presents a summary screen with reps completed, visualized progress, and performance metrics.

How we built it

GymBro AI is built using Python with a Streamlit frontend. We use pose estimation to analyze body movement and detect state changes during exercises such as push-ups and tricep dips.

The system integrates:

DigitalOcean Web Deployment for hosting the app
DigitalOcean Agent Platform to power the GymBro AI personality
AI generated coaching responses during workouts
Text to Speech (TTS) so GymBro can talk back to the user
Session state management to control workout flow and UI transitions

This reflects our original vision of a fully connected AI-powered workout experience.

Challenges we ran into

Initially, connecting DigitalOcean services to our codebase felt intimidating. However, once we learned the platform, it became one of the strongest parts of our project. The agent system, deployment tools, and infrastructure allowed us to implement features like AI coaching and hosting much faster than expected.

Another challenge was maintaining a smooth live camera feed while continuously updating the UI and rep counter without causing flicker or instability.

Accomplishments that we are proud of

We are especially proud of building a relatively accurate pose estimation and rep counting system, especially given the dark and chaotic hackathon environment. Getting real-time computer vision working reliably as freshmen was a major accomplishment for us.

What we learned

We learned how to design and build a full-stack application from scratch. This included integrating computer vision, AI agents, cloud deployment, UI state management, and running Python as a live web application.

As freshmen, this was our first time building something at this scale, and we learned an incredible amount in a very short time.