Prudhvi Kadamuthuri

Data & AI Engineer

I build ML systems that move from model training to real product experiences. This project demonstrates production-style AI engineering: data pipeline, model training, API serving, and user-facing integration.

---

Neural Collaborative Filtering Recommender System

What it solves

Helps users discover relevant movies faster by ranking items based on historical user behavior.

What I built

• Designed a MovieLens 100K data pipeline for ratings and item metadata
• Implemented a Neural Collaborative Filtering model in PyTorch
• Built the full training + evaluation workflow with ranking metrics
• Implemented a FastAPI inference backend (/health, /recommend)
• Integrated recommendations into a frontend experience (with this repo also containing a Streamlit demo app)

Tech stack

• Python
• PyTorch
• FastAPI
• Next.js

Output

• Training pipeline: data preprocessing, negative sampling, model training, evaluation
• API layer: production-style inference endpoints
• UI layer: interactive recommender demo

---

Architecture Summary

User → API → Model → Recommendations

1. User submits user_id
2. FastAPI receives request at /recommend
3. Trained NCF model scores candidate items
4. Top recommendations are returned

---

Live System (if available)

• Render Backend API: Deploying — Add Render URL
• Frontend UI (Vercel): Deploying — Add Vercel URL

---

Key Achievements

• HR@10 ~ 0.54
• NDCG@10 ~ 0.31
• Production-style API for inference
• Artifact persistence for serving (ncf.pt, user2idx.json, item2idx.json, items.csv)

---

Key Skills Demonstrated

• ML systems design
• PyTorch training pipeline development
• FastAPI backend engineering
• End-to-end deployment workflow

---

Run Locally

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python3 -m src.train
uvicorn src.service.api:app --reload
streamlit run app/app.py