Projects
Production-focused ML systems. Built to solve real problems, not to exist on GitHub.
Advanced RAG System v2
Upgraded a basic RAG pipeline into a production-grade system using patterns from Netflix and Amazon: hybrid BM25 + dense retrieval with Reciprocal Rank Fusion, LLM query rewriting (+15-25% recall), cross-encoder reranking, and token-by-token streaming. Evaluated with RAGAS across faithfulness, relevancy, and context precision.
RAG Knowledge Assistant
A production-grade Retrieval-Augmented Generation system built with LangChain, FAISS, and HuggingFace Sentence Transformers. Ingests multi-format documents (PDF, DOCX, web), chunks intelligently, embeds to a local vector store, and answers questions with source citations. Supports swappable LLM backends (Ollama, OpenAI, Groq).
Diabetic Retinopathy Detection
Deep learning pipeline for automated diabetic retinopathy severity grading from retinal fundus images. Uses a fine-tuned EfficientNet backbone with custom augmentation strategy for handling class imbalance in medical imaging. Achieves clinician-grade accuracy on the APTOS benchmark.
Insurance Claims Processing (GCP)
End-to-end ML pipeline for automated insurance claims classification and routing on Google Cloud Platform. Combines Document AI for structured extraction, BERT-based NLP for intent classification, and BigQuery ML for risk scoring. Deployed on GKE with KubeFlow orchestration and real-time monitoring.