RepRight

Inspiration

Getting started with fitness can be hard. You want to build consistency but sometimes you aren't even quite sure if you’re doing an exercise correctly. Proper form is critical for both progress and injury prevention, yet it can be surprisingly inaccessible in many cases. In a gym, asking others to check your form can sometimes feel awkward or inconvenient, and not everyone has access to a personal trainer. Online workout videos help, but they’re passive. You can try to mirror a video, but you probably still won't be 100% confident that your posture, range of motion, and joint alignment are correct. We felt there should be an easier, more immediate way to get feedback. In particular we wanted to create a way that doesn’t require another person, special equipment, or stopping your workout to rewatch tutorials. That idea led us to build RepRight.

What it does

RepRight is an AI-powered exercise form coach that provides real-time feedback from your video feed. You simply turn on your camera and start your exercise. As you move, the system uses computer vision to track key body points and measure important aspects of your form such as posture, joint alignment, and range of motion. If your form begins to break down, you receive immediate, on-screen alerts so you can correct it during your set. After completing a set, you can click “Get coach feedback” to receive a clear summary of feedback, highlighting what you did well and what needs improvement. Users can also create an account to save their workout history. This allows you to track your progress over time, view past feedback, and see how your form consistency improves across sessions.

How we built it

RepRight is architected to support a growing inventory of exercises, users, and fitness coaching features.

Core tech stack: Our core functionality uses mediapipe as our CV solution for robust detection of body positioning and movement. We use GPT-4 through backboard.io to provide a coaching summary with integrated memory persistency.
User accounts: We use Auth0 to handle secure user authentication and account management. Auth0 allows users to securely sign up and log in while keeping authentication logic separate from our core application, enabling safe access to personalized data and future identity provider support.
Session history & persistence: We use MongoDB to store user sessions, exercise metadata, and coaching feedback. This allows users to view their full history in one place, track form improvements over time, and enables us to flexibly evolve our data model as we add new exercises and coaching features.

Challenges we ran into

Linking body tracking with exercise-specific expectations: While MediaPipe provides accurate body keypoints, translating raw joint positions into meaningful form checks for different exercises was a challenge. Each movement has unique constraints, ranges of motion, and failure modes, requiring us to design logic that adapts body tracking data into exercise-aware feedback.
Balancing real-time feedback with post-set coaching: Having both real-time flags during a set while also generating higher-level coaching summaries introduced performance and design challenges. We needed to ensure real-time analysis remained fast and non-intrusive, while still capturing enough context to produce thoughtful, actionable feedback once a set was complete.
Designing for scalability across exercises: We wanted to focus on building an architecture that could easily expand beyond a small set of movements. But the initial refining of our models and feedback logic required large time investments into few exercises.

Accomplishments that we're proud of

Meaningful and actionable feedback: We’re proud that RepRight goes beyond simply detecting movement and instead delivers feedback users can actually apply. By translating complex body tracking data into clear, practical coaching cues, the tool allows users to have a high degree of confidence in the corrections being made to their form.
Easy of use: We’re especially proud of how simple the experience of RepRight is. Our solution requires no special equipment or setup as users can simply just turn on their camera and start correcting their form. Keeping the interface intuitive and unobtrusive ensures that feedback supports the fitness process instead of making things needlessly complicated.

What we learned

We learned that real-time computer vision requires architectural trade-offs. In particular, we found out that delivering real-time form feedback isn’t just a computer vision problem, but one that shapes the entire platform design. To keep latency low while maintaining real-time feedback, we had to carefully decide which computations belonged in the live feedback loop and which were better handled after a set was completed.

What's next for RepRight?

A larger inventory of exercises: Our current application is easily adaptable for a endless database of exercises to choose for from. However, the actual implementation would need to ensure high standards of accuracy and would be a great thing to work in the future.
Featureful AI fitness coaching: We think the RepRight platform can be developed to have a full suite of fitness coaching features. There could be a complete AI fitness coach with suggestions fully customized based on the needs and goals of every individual user.