Designing a Clinical Evidence Operating System – How Domain-Driven Agentic Decomposition Minimizes Hallucination Without Scaling Model Size

Abstract:

Hallucination remains the central barrier to deploying large language models in evidence-based medicine. The dominant response has been to scale model size. We argue this is the wrong lever. This talk presents Open Prometheus, a live clinical evidence platform built on a governed, multi-stage agentic pipeline. We show that domain knowledge encoded into workflow design, not raw model capacity, is the true lever for reliability and trust in clinical AI. The talk concludes with practical lessons on building trustworthy clinical AI under EU AI Act and GDPR constraints, and open questions on evaluation methodology for agentic medical systems.

About André:

AbdulRahman Ulfat is a physician and AI researcher at Universitätsklinikum Hamburg-Eppendorf (UKE), where he conducts deep learning research on fracture detection in X-ray imaging. Prior to this, he worked in sports medicine. He is the founder of Open Prometheus, an EU AI Act-compliant clinical evidence retrieval platform providing access to 35M+ scientific publications, live since November 2025. He is an alumnus of the Antler Accelerator Berlin.