EHSAN Imani

I designed and conducted a physician-reviewed benchmark of Qwen3:8B using 20 original USMLE-style clinical cases. I evaluated its answers with a structured physician rubric and ran a separate automated evaluation through the local Ollama API.

In this pilot, the model scored 152/160 on the physician rubric, 18/20 on the independent answer-choice run, and 14/15 on a perturbation test. The evaluation also surfaced a clinically important safety error in a DKA case, showing me why answer accuracy alone isn't enough.

At Shahid Sadoughi University of Medical Sciences and Health Services, I conducted clinical literature reviews and evidence synthesis across medical research projects. I contributed to clinical data evaluation, research methodology, and manuscript preparation, including work in orthopedics and women’s health.

I'm interested in how medical AI performs on clinically relevant tasks, particularly its accuracy, safety, reasoning, and robustness. I want to keep learning, build practical evaluation projects, and explore opportunities in Medical AI and clinical research.