Lightweight retrieval refinement for RAG systems

• Retriever-agnostic refinement library for RAG and agentic AI workflows
• Implemented neural, lexical, and pattern-based ranking with deterministic RRF, duplicate suppression, and Top-K/token budget context selection, giving agents smaller, higher-signal context with auditable traces.
• Built reproducible BEIR/Ragas evaluation

Visual explanations for CNN predictions using Grad-CAM

• FastAPI + PyTorch + ResNet18
• Grad-CAM heatmap visualization
• Explainable AI

Voxel-wise segmentation of the tumor mapping distinct tumor subregions including whole tumor, tumor core, and enhancing tumor regions.

• Build a 3D-UNET architecture to perform brain tumor segmentation on multi-modal MRI scans (BraTS dataset), processing T1ce, T2, and FLAIR imaging modalities.
• Fully modularized pipeline with separate, reusable components for preprocessing, dataset handling, model definition, training, evaluation, and visualization.

Implementation from transformer architecture using Pytorch

• Built transformer architecture from scratch using PyTorch with complete encoder-decoder implementation.
• Trained and validated sequence-to-sequence translation model

ML/AI: PyTorch · Hugging Face · LangChain
Backend: FastAPI · Flask · Streamlit
DevOps: Docker · CI/CD · Azure
Data: SQL · Vector Databases (Pinecone, ChromaDB)