Soumyadeep Roy

I work on deployable medical language models — making them efficient enough to run in real clinical workflows, reliable enough for clinical decision-making, and auditable against established medical guidelines.

I am a postdoctoral scholar working with Prof. Tina Hernandez-Boussard at the Division of Computational Medicine, Department of Medicine of Stanford University. My current research at Stanford focuses on the clinical domain of perioperative pain management where I collaborate with a clinical team of surgeons, trauma care physicians and psychologists. I am leading the following two research directions - (i) developing auditable reasoning LLMs for clinical tasks, using clinical guidelines as the evidence base, and (ii) LLM as a tool to construct executable guideline pathways and understand deviations in real-world patient trajectories, to improve patient care outcomes.

My research spans three connected threads: efficient pretraining and adaptation of medical and biomedical foundation models (GeneMask, vocabulary adaptation during finetuning, adaptive BPE); reliability evaluation of clinical LLMs (USMLE error taxonomy, OOV impact studies, trustworthy AI tutorials); and guideline-grounded reasoning for clinical decision support. Published 16+ peer-reviewed papers at SIGIR, ACL, IJCAI, EMNLP, ECAI, CIKM, and Frontiers in AI.

Please go through my resume for further details.

Previously: PhD from IIT Kharagpur (2025), Research Associate at L3S Research Center / Leibniz University Hannover (2021–2023), AI Intern at Wipro GE Healthcare (2024–2025), and Research Intern at Adobe Research (2018).