GenFit AI

• 

  logo

Inspiration

Physical health is a universal necessity, yet access to high-quality personal training remains a luxury reserved for the elite. We observed a growing physical inactivity epidemic, particularly among the youth who are increasingly tied to their screens. We asked ourselves: > "How can we democratize elite-level personal training and biomechanical tracking using hardware everyone already owns?"

The inspiration for GenFit AI emerged from the desire to create a "Green AI" ecosystem. Traditional AI fitness apps rely on expensive cloud GPUs for processing, leading to massive subscription costs and a large datacenter carbon footprint. By engineering an intelligent Edge-Computing architecture, we realized we could process everything natively on the user's device.

What it does

GenFit AI is a comprehensive, cross-platform hybrid ecosystem offering:

• The Edge Neural Posture Coach: Automatically counts your reps and analyzes your joint-angle tracking in real-time (Squats, Pushups, Lunges) entirely via your smartphone's front camera. It provides immediate Voice-Actuated form correction.
• FitBot LLM Assistant: An agentic AI built into the app that can dynamically generate tailored diets and weekly training vectors specifically matched to user macros.
• Intelligent Diet Generation: Seamless, auto-generated metabolic meal plans focusing on Hypertrophy, Endurance, and Weight Loss.
• Global Admin Dashboard: A unified full-stack web ecosystem for community management and engagement metrics.

How we built it

We aligned the architectural choices of GenFit AI explicitly with the FOCUSS Framework (Financial Viability, Originality, Customer Applicability, Usefulness, Sustainability, Structured Approach).

We built the core biomechanical tracking by capturing coordinates from the TensorFlow MoveNet neural engine. To calculate the 3D joint extension (for example, the knee angle during a squat), we used vector mathematics directly on the Edge hardware.

We calculate the joint angle \(\theta\) using the dot product formula of the adjacent bone vectors \(\vec{A}\) and \(\vec{B}\):

$$ \theta = \arccos\left(\frac{\vec{A} \cdot \vec{B}}{|\vec{A}||\vec{B}|}\right) \times \frac{180}{\pi} $$

If the knee angle \(\theta < 100^{\circ}\), the system recognizes a valid squat depth and logs the rep asynchronously:

// Localized Edge Logging
if (angle < 100.0) {
  this.state = 'down';
} else if (angle > 160.0 && this.state === 'down') {
  this.reps++;
  this.state = 'up';
}

Challenges we ran into

The most severe hurdle was bypassing memory exceptions thrown by Native GPU architecture. Operating a 60fps camera feed and feeding those frames directly into a WebGL-backed Neural Engine inside a React Native container caused repeated memory kernel panics (SIGSEGV) on low-end Android hardware.

We overcame this by architecturally isolating the tracking engine from the main React ScrollView view-tree, creating a segregated off-screen canvas buffer to parse image tensor data asynchronously. This breakthrough allowed us to retain true real-time skeletal mapping without sacrificing device stability.

Accomplishments that we're proud of

What we learned

We learned extensive details regarding React Native rendering life-cycles, the strict limitations of Android GPU memory management, and the vast mathematical challenges of applying cosine-angle vectors across dynamic 2D coordinates to accurately measure biomechanical kinetics.

What's next for GenFit AI

We plan to introduce deeper gamification (AR badges, competitive rep-leaderboards) and expand our tracking library to accommodate complex Yoga asanas and dynamic pliometric movements, ultimately launching the application on the Google Play Store as the ultimate free fitness companion for India and the World.

Built With

• express.js
• mongodb
• movenet
• node.js
• react.js
• tailwind-css
• tensorflow.js