At Visdom Lab, IISER Bhopal, I developed a medical image segmentation framework using CLIP with attention-based adapters. It achieved 86% accuracy on LIDC-IDRI and an 88% Dice score on NPC-120.
I also implemented aleatoric and epistemic uncertainty estimation with probabilistic deep learning and Monte Carlo sampling, and developed a learnable framework to model annotation variability across expert annotators.
For a document-based RAG chatbot project, I developed PDF and Markdown ingestion with context-aware question answering. I built its retrieval pipeline using Hugging Face embeddings and LangChain, with SQLite for conversation history.
In deep learning projects, I aligned Qwen3-0.6B-Base on GSM8K using LoRA-SFT and GRPO, improving reasoning accuracy from 49.13% to 55.37% while training 0.84% of parameters. I also implemented an autograd engine from scratch and trained a Vision Transformer on CIFAR-100.
