Designed end-to-end CNN-based medical imaging pipelines with advanced denoising techniques and applied multiple explainable AI methods (Grad-CAM, Score-CAM, Ablation-CAM) for interpretable predictions. Published and presented research at NeurIPS workshops and MDPI journals, with a focus on robust medical image analysis and XAI.
This work presents a deep-learning-based framework for visualizing and estimating retinal fluid volumes in OCT scans across IRF, SRF, and PED pathologies. Using Inception-ResNet-based models and a robust Ensemble-CAM visualization approach, the method enables interpretable localization and accurate volumetric analysis validated against expert annotations.
Built an end-to-end Retrieval-Augmented Generation (RAG) pipeline for document-grounded question answering using SQuAD v2 as a benchmark. Designed a structure-aware chunking strategy with controlled overlap, indexed document chunks in Chroma using MiniLM bi-encoder embeddings, and evaluated retrieval quality using Recall@K based on gold answer spans. Enhanced precision by introducing a cross-encoder reranker to re-score candidate passages, significantly improving fact-level retrieval accuracy. The system mirrors industry-standard search pipelines by separating fast candidate recall from high-precision relevance ranking.