Logo of FitBridge
Home Page

Inspiration

As a long-time gymer, I have deeply understood the struggles that beginners face: the complete lack of guidance to correct their lifting form and the difficulty of tracking progress effectively. To make matters worse, many of my friends have spent a fortune hiring personal trainers, only to end up with inexperienced or outright scam trainers who don't deliver results. We wanted to build a solution that democratizes high-quality AI fitness coaching (making it accurate, affordable, and accessible to everyone). Track: Health for All

What it does

FitBridge is an AI-powered virtual fitness assistant that turns your camera into a real-time personal trainer. Using advanced computer vision, the app tracks your body movements during an exercise and provides instant, real-time audio feedback to correct your form using LLM and AI-generated text-to-speech. At the end of every set, FitBridge generates a comprehensive analytical report displaying critical workout metrics: total calories burned, exact rep/set counts, and an automated form audit highlighting specific posture mistakes to fix for the next set.

How we built it

We built FitBridge with a robust, modern tech stack optimized for real-time processing and intelligent feedback:

Frontend & Deployment: Developed a responsive user interface deployed seamlessly on Vercel.
Computer Vision Core: Integrated OpenCV and MediaPipe Pose for high-performance, low-latency 2D/3D human body skeletal tracking directly via the browser camera. Made a joint angle benchmark for a qualified rep.
AI & Voice Engine: Leveraged the Google Gemini API to instantly analyze joint angle data and generate contextual form corrections, which are then converted into real-time audio feedback using a Text-to-Speech (TTS) pipeline.

Challenges we ran into

Model Heavyweight & Latency: We initially attempted to use YOLOv8 for body tracking, but it proved to be too heavy and caused severe lag when deployed in a real-time web environment on Vercel. To fix this, we pivoted to a lightweight setup using MediaPipe combined with OpenCV, which drastically reduced latency.
False Rep Counting: In the early stages, the computer vision accuracy for tracking reps was low-even a simple nod of the head would mistakenly trigger a rep count. We resolved this by completely refactoring our counting logic, establishing strict geometric rules, and setting precise joint angle thresholds to validate a complete range of motion.

Accomplishments that we're proud of

High-Accuracy Tracking: After rewriting our rep-counting algorithms, we significantly boosted the precision of our computer vision engine, minimizing false triggers.
The Goggins Factor: We successfully integrated an aggressive, highly motivating AI audio persona inspired by David Goggins-the ultimate idol for anyone pushing past their limits-to drive users to smash their goals.

What we learned

Through this intense development journey, our team grew immensely. We mastered efficient time management under pressure, optimized our workflow for seamless team collaboration using Git/GitHub, and perfected our prompt engineering skills with Claude to generate highly accurate structural responses from our AI engines.

What's next for FitBridge

Moving forward, we plan to scale FitBridge into a comprehensive fitness ecosystem by dramatically expanding our exercise library to support complex compound lifts and targeted routines. Ultimately, we want to transform FitBridge into a thriving, community-driven social platform where fitness enthusiasts can connect, share their progress charts, and push each other to stay uncommon.

Built With

css
google-gemini-api
javascript
mediapipe
next.js
opencv
plpgsql
postgresql
react
supabase
text-to-speech
vercel

Submitted to

Global Builders Hackathon: Code for Impact

Created by

I was responsible for building the core computer vision engine and managing the end-to-end repository integration. I integrated OpenCV and MediaPipe Pose to handle real-time skeletal joint tracking, and developed a robust geometric rule-based counting algorithm to ensure highly accurate rep detection. Additionally, I set up the backend database structures with Supabase (using PLpgSQL) and managed our team's version control on GitHub to ensure clean, error-free production code.

gia khánh tạ
I served as the Business and Finance Strategist for FitBridge. I conducted extensive market research and competitive analysis within the fitness training industry to validate our product-market fit against expensive traditional personal trainers. I designed our monetization tiers, analyzed the financial feasibility and cloud hosting costs for our AI infrastructure, and built the project's financial model. Additionally, I took the lead in crafting the pitch deck to ensure our business metrics and value proposition are clearly presented to the judges.

Andy Tran