Hoop Master

AI-assisted basketball training — pose-based analysis of shooting and drill form.

Role: Two-person final year project. I built the computer-vision and Python backend: pose analysis, form scoring and feedback. My teammate built the mobile front end and the coaching chatbot.
Period: Final year project · completed July 2025
Status: FinishedNot hosted

Context

Why it exists

Personal basketball coaching is expensive and scarce, and the apps filling that gap mostly ship a drill library. Hoop Master was our final year project at COMSATS University Islamabad: a player records a drill or a shooting rep and gets back a scored assessment of their form.

Architecture

How it is put together

A layered stack that keeps the AI services isolated behind the domain API. External integrations are reachable only through the application tiers. I owned the AI services layer and the Python backend under it.

01
Presentation
Built by my teammate
- Mobile app (React Native)
- Admin panel (Next.js)
- Camera capture
- Notifications
02
Domain services
- Authentication
- User management
- Course management
- Social hub
- Subscriptions
03
AI services
My work, except the chatbot
- Drill form analysis
- Shooting practice analytics
- Chatbot query handling
04
Data
- MongoDB
- Analytics and reporting
05
External integrations
- Cohere (NLP)
- Wise (payments)
- Nodemailer (email)

Decisions

Trade-offs

Keep pose analysis in its own Flask service behind the Node/Express API
Instead of Run the analysis inside the Node API, or on the mobile client
The mobile client never calls the model, and the Python CV work could be built and deployed on its own. The cost is a second runtime to operate and an extra hop on every analysis request.
Derive form metrics from the landmark sequence over time
Instead of Score individual frames and average the result
A rep is judged as a movement rather than as a pose. The cost is that a noisy or dropped stretch of frames corrupts the whole attempt instead of one frame of it.
Return an annotated video with the score
Instead of Return the score and the written corrections alone
The player can see the moment in the rep the feedback refers to. The cost is that every analysis has to render and store a video file, not just produce a number.

Constraints

Limits and how they are handled

Analysis runs on an uploaded recording, not a live camera feed.: The player records a rep in the app; the Flask service returns the score and annotated video afterwards.

Media