At IEEE Computer Society Bangalore, I built a federated training pipeline for medical image segmentation with Flower and PyTorch. I compared FedAvg and FedProx for SWIN UNETR under non-IID data partitions and added DP-SGD with Opacus to study the privacy-utility tradeoff.
For Aero-Astra, I built a multi-agent system for satellite fault management, connecting a FastAPI backend and React/Vite frontend to a WebSocket telemetry stream. I designed SENTINEL’s hybrid anomaly detector and evaluated it on OPS-SAT and Mars Express telemetry.
I led a team building Serenity Bot, a campus mental well-being chatbot. After identifying retrieval quality gaps, we moved from Naive RAG to Modular RAG; our evaluation findings are drafted into a research paper, and the deployed system achieved a 40% latency reduction compared with traditional search methods.
I’ve also worked on semantic code search using AST-based chunking and RAG, and explored dynamic A* for satellite anomaly recovery. I’m interested in RAG systems, federated learning, and privacy-preserving machine learning.

