Mahesh Keswani

I'm pursuing an MSR in Artificial Intelligence at IIT Delhi, where my research focuses on safe and risk-sensitive reinforcement learning.

I developed offline RL methods for safe decision-making under distribution shifts, achieving 45% higher performance and 17× faster inference than the strongest baseline. My work also includes uncertainty-aware SAC and safety-constrained learning under uncertain environment dynamics.

At LTIMindtree, I architected an Oracle GoldenGate migration framework that saved $42.69K annually and built a RAG-based legal chatbot backend with AWS Kendra that reduced information retrieval time by 94%. I also contributed to a centralized logging pipeline processing 288GB of distributed application logs per day.

I've worked across multilingual knowledge distillation, cross-lingual relation extraction, autonomous-driving safety benchmarking, and RLHF-based research-statement generation. As a graduate teaching assistant, I develop assignments and conduct viva-voce examinations for graduate reinforcement learning and machine learning courses.