At IIT Bhubaneswar, I built ARA-FL, an asynchronous federated LLM fine-tuning framework that uses a Double DQN agent to select LoRA rank and QLoRA bit-width. It cut average VRAM by up to 70% and cross-entropy loss by 27% compared with FedAvg.
I also built and deployed a legal-document RAG chatbot using FAISS, Streamlit, Anthropic Claude, and OpenAI. In a separate object detection and retrieval project, I created a CPU-only pipeline using YOLOv11s, BoT-SORT, CLIP, and FAISS to find objects’ last-seen times and frames.

