Production-style retrieval-augmented generation system for document Q&A. Chunks and embeds PDFs using OpenAI embeddings, stores in FAISS, and serves answers via a FastAPI endpoint with citation tracking and query reranking.

Fine-tuned Llama 2 on domain-specific instruction data using QLoRA and PEFT. Reduced GPU memory footprint by 60% while matching full fine-tune performance. Tracked experiments with MLflow and deployed via SageMaker.

Full-stack expense-splitting web app built with Next.js, Supabase, and Prisma. Supports equal splits, exact amounts, percentages, and shares — with real-time balance tracking.

End-to-end ML serving system with experiment tracking, model registry, and versioned REST API endpoints. Containerized with Docker, orchestrated on Kubernetes, with automated retraining triggers on data drift.

Credit risk models using XGBoost and Neural Networks on the Amex dataset. Engineered temporal features, tuned hyperparameters, and used SHAP for model interpretability.

BERT-based named entity recognition and sentiment analysis service deployed with FastAPI and Docker. Handles multi-label text classification, entity extraction, and aspect-based sentiment across customer feedback data.