Kinera Flow

Inspiration

Fitness is essential for both physical and mental well-being, yet many people struggle with proper form and technique. Beginners may feel intimidated by the gym, while experienced athletes often lack real-time feedback to refine their movements. Without guidance, users risk injury and fail to get the most out of their workouts.

We wanted to build an accessible, affordable alternative to personal training that provides real-time coaching using just a computer and a webcam. Existing real-time form analysis solutions cost thousands of dollars. Ours does not.

What it does

During the workout

Uses live computer vision and motion capture to track body movement
Analyzes exercise form in real time against proper technique standards
Provides pacing guidance to help maintain consistent tempo

After the workout

Highlights form issues and areas for improvement
Suggests exercise-specific adjustments tailored to the user

How we built it

We use MediaPipe for real-time pose estimation due to its speed, reliability, and limb modeling capabilities. The frontend UI is built with Tauri and Vite, while Rust handles workout selection and real-time metric display.

Exercise analysis and scoring are handled by a custom PyTorch-based model, which processes numerical data derived from joint angles, vectors, and timing. These results are condensed into decision logic that evaluates form quality and generates feedback.

Challenges we ran into

One of our biggest challenges was training the AI model without overfitting to limited data. We also encountered difficulties with 3D vector math, especially when users were positioned at an angle relative to the camera, which could distort joint angles.

To address this, we implemented a hybrid approach: most joints are analyzed in 2D for stability, while certain limbs switch to 3D calculations when severe camera angles are detected. We also built a video upload pipeline that allows additional training using human-reviewed weighting to improve accuracy over time.

Accomplishments we’re proud of

A fast, clean, Material 3–inspired UI that remains responsive despite heavy real-time processing
Highly accurate computer vision and vector math capable of detecting reps with over 95% accuracy across different speeds and body types
A custom quantitative AI model that evaluates form with strong precision

What we learned

Training AI on numerical data is surprisingly difficult. Models cannot “see” graphs or movement intuitively, so every insight must be encoded numerically. Weighting this data introduces human bias, and exercise form itself is inherently subjective. There is rarely one perfectly correct way to perform a movement, which made defining success metrics especially challenging.

What’s next for Kinera Flow

With more time, we would port Kinera Flow to Android for greater accessibility and explore hardware-assisted depth sensing for improved 3D accuracy. We would also expand the exercise library significantly.

Thanks to our modular design, adding new workouts or improving analysis requires minimal additional engineering. The foundation is already there!