Jared Joselowitz is a senior AI research engineer at Ufonia, developing safety and evaluation systems for Dora, a conversational voice agent that handles clinical follow-ups and preoperative assessments. His work addresses an unusually strict engineering constraint: patients cannot ethically be exposed to unsafe experiments, and harmful medical advice cannot be recalled after a conversation ends.
Originally from South Africa, Joselowitz studied electrical engineering at the University of the Witwatersrand, worked as a software engineer at Tesserae and a data scientist at The Awareness Company, and pursued graduate research at Imperial College London before joining Ufonia in 2024. His work on inverse reinforcement learning for language models investigated whether observing model behavior can reveal learned reward functions, illuminate biases, and expose alignment failures. A fundamental complication is that different objectives can produce similar behavior.
Making clinical AI accountable
- Hazard-driven patient simulation. Joselowitz co-authored MATRIX, which tests clinical conversational systems against hazards including missed warning signs, fabricated medical guidance, and ignored patient distress. A simulated patient generates clinically grounded exchanges; an automated evaluator flags dangerous behavior against expectations checked by clinicians. Patient and public involvement helps assess realism across different communication styles.
- Clinical risk-weighted optimization. Missing a serious symptom is more dangerous than asking unnecessary follow-up questions. Joselowitz uses clinician-informed evaluation criteria and automated prompt optimization to prioritize detecting red flags, while preserving auditable links between identified hazards, prompt versions, and evaluation results.
- Safer retrieval and transcription. He co-authored ASTRID, a safety framework for retrieval-augmented clinical question answering, and research on clinically consequential speech-recognition errors. The latter challenges conventional word error rate: a single mistranscribed symptom can matter more than numerous medically inconsequential wording differences.
- Evidence-gated clinical deployment. Joselowitz’s AI Engineer World’s Fair presentation describes simulation as preparation for supervised clinical evaluation, monitored deployment, and increased autonomy only as supporting evidence accumulates. Voice interactions introduce additional hazards, including interruptions that prevent patients from receiving complete safety advice. He has also shared this patient-facing evaluation approach with colleagues at The Alan Turing Institute.
