1) Built an Agentic RAG assistant (LangGraph, LangChain, ChromaDB) used by 500+ radiology staff to select CT Imaging protocols

2) Optimized inference with vLLM batching, Asyncio, reducing latency by 40% for secure deployment in the private hospital network

3) Deployed as a FastAPI service with Docker. Reduced protocol lookup time by 60% through direct answers with source references

1) Collaborated with Flywheel engineers to extract, preprocess 80K+ unstructured radiology reports for downstream analysis

2) Built an Agentic RAG system (LangGraph + Langchain + LLaMA) for medical QA, evaluated on 1K+ reports using ROUGE metric

3) Dockerized and Deployed as Flywheel Gear reducing lookup time by 70% and forming the base for the client’s multimodal RAG

1) Tackled the resource-intensive task of labeling datasets and fine-tuning LLMs by designing Uncertainty-based and K-Center Clustering data selection strategies to identify the most informative prompts

2) Fine-tuned LLaMA-2 on CUDA GPUs, achieving 53% win rate (AlpacaEval), reducing training data by 67% & GPU runtime cost by 51.4%

3) Demonstrated accuracy comparable to full dataset fine-tuning, showing 3x efficiency gains without loss in model performance

4) Validated results through statistical hypothesis testing and monitored training with Weights & Biases to ensure reliability