Building in Explainable AI space for critical domains that keeps AI and humans using it in control.

Hi! I'm a Research Engineer at JazzX AI, where I work on reward modelling, human-in-the-loop design, formal methods, and model evals and benchmarks. Right now I build AI-powered solutions with financial practitioners in mortgage to accelerate the process of document reviewing for loan approvals.

My concern with the AI is a slow one, where people review less, skim more, and lose the critical thinking leading to oversight and lose of control. I have watched this happen with doctors and with financial experts. I think humans and AI should co-evolve: AI should make people better at their work, and people should keep real control over what AI does. This shapes my focus on AI control, human oversight, and gradual disempowerment.

Previously, I was a Research Fellow at Microsoft Research India. I built CataractBot, an expert-in-the-loop chatbot where doctors verify answers to cataract patients, and led AshaBot, in collaboration with KhushiBaby NGO, a chatbot for community health workers. I also studied how doctors review LLM drafts. I have also worked on LLM inference, writing custom CUDA kernels for KV-cache memory management in vLLM.

Before research, I worked as a software engineer at Microsoft on Teams and at Rubrik on cloud backup and recovery.

If you work on AI control, human oversight, or evals for high-stakes domains, I'd love to talk. Drop me a message!