I build LLM retrieval systems, ML models, and the APIs that serve them. Two are live: a RAG pipeline at 0.955 hit rate, and an ML recommender at 0.97+ R². Both measured, both deployed.
A point of sale and stock system for a hardware store. Django and DRF over PostgreSQL, React on the counter. Built around one rule: stock can never be oversold, even when two cashiers bill the last unit at the same instant.
A 3-stage retrieval pipeline over nine Indian labour-law documents. BM25 keyword search fused with vector semantics by Reciprocal Rank Fusion, then cross-encoder reranking down to five passages. Quality gates block deploys when metrics slip.
Ranks 25 packaging materials across 13 product categories by suitability, cost and carbon impact. A composite eco-score makes the ranking readable; eleven REST endpoints and a normalised PostgreSQL schema make it usable.
The patient-facing side of a hospital system: browse doctors, book appointments, register, and a protected dashboard. Refactored out of a duplicated HTML/Bootstrap site into a React component architecture with custom hooks for fetching and debounced search. Backend is next.
Python · SQL · JavaScript · Data Structures & Algorithms · OOP · REST API design · System design basics · Statistics & probability
LangChain · Google Gemini · HuggingFace Transformers · ChromaDB · Langfuse · Hybrid search (BM25 + vector) · Cross-encoder reranking · Retrieval evaluation
PyTorch · scikit-learn · XGBoost · Pandas · NumPy · Feature engineering · Tableau
Django · Django REST Framework · FastAPI · Flask · JWT auth · Pydantic · Uvicorn · Middleware & rate limiting · Transactions & row locking · Pytest
React · Vite · Next.js · TypeScript · Tailwind CSS · Streamlit
PostgreSQL · MySQL · SQLAlchemy ORM · Schema design · Query optimisation · ChromaDB
Docker · Docker Compose · Git · GitHub Actions · CI/CD · Google Cloud Run · Render · Vercel · Linux
A retrieval system that answers well in a notebook and drifts in production isn't finished. So I build the evaluation harness first: a 22-question dataset, LLM-as-judge scoring for faithfulness and citation accuracy, and quality gates in CI that refuse the deploy when the metric degrades.
Six months of paid analyst work taught me the other half: the model is only as useful as the question it answers. Before that, a YOLOv8 terrain-identification paper and a traffic-sign detector at 92% accuracy. Now: Hyderabad, and open to AI and full stack engineering roles.
Final year project: real-time traffic sign detection with YOLOv8, 92% classification accuracy.