Urbana, Illinois, United States
Machine Learning Engineer with ~4 years of experience in NLP, Generative AI, and Computer Vision. UIUC graduate (4.0 GPA), working with Profs. Gokhan Tur, Dilek Hakkani-Tur & Suma Bhat on LLM healthcare agents, function-calling systems, and large-scale multi-GPU training with Accelerate + FSDP. At Microsoft, enhanced multi-turn conversational AI with GPT-4 & SLMs, optimized intent detection with CoT prompting, and improved Azure AI Search retrieval-boosting F1 and reducing latency. Prior work at Novelis applying GANs and GenAI to SEM image segmentation & alloy optimization; and at VerSe Innovation scaling CV + retrieval systems for 200M+ users using Vision Transformers, Neo4j, and video captioning. Strong MLOps and research background; experienced in building robust RAG, RLHF, and multi-modal AI systems. Tech Stack: Python, PyTorch, TensorFlow, LangChain, FSDP, Docker, AWS, GCP, RAG, LLMOps. Actively seeking collaborations in responsible AI and real-world ML. Reach out: [email protected]
Developed advanced multi-turn, multi-intent conversational AI Agents beyond Azure CLU using GPT-4 and fine-tuned Phi-3.5 Mini models for intent detection and entity recognition. Built synthetic data pipelines , improved F1 score by 18% while reducing latency. reducing SGD error rates below 3% and enhancing RAG-based enterprise response systems.
Restructured dialogue state data and fine-tuned LLMs for function-calling agents in task-oriented dialogues for conversational AI, under the mentorship of Professors Gokhan Tur and Dilek Hakkani-Tur. Enhanced patient note grading capabilities using Chain-of-Thought and React techniques guided by Professor Suma Bhat. Leveraged Reinforcement Learning within REALM frameworks for question-answering tasks, optimized LLM training speed with DeepSpeed, Accelerate, and FSDP, and conducted extensive experiments on PEFT-based fine-tuning for NLU tasks. Developed RAG solutions for medical note grading and explored training efficiencies using Direct Preference Optimization.
Led image segmentation on microstructural alloy SEM images using UNet and applied Generative AI in materials science to train GAN models with segmented images. Fine-tuned Llama3 for structuring alloy production data, enhancing data usability and analysis, and collaborated with cross-functional teams to streamline data collection from non-technical personnel.
➣ Presented this work at the 7th AI and Health Summit, Chicago. Developed CNNs to differentiate cognitively normal from those with dementia using fMRI scans, achieving 96% accuracy. ➣ Enhanced MRI data accuracy through advanced pre-processing techniques like registration and segmentation, and streamlined data storage and model training on AWS using pyspark and CI/CD pipelines