Summary
Jared Joselowitz, a research engineer at Ufonia, presents at the AI Engineer conference on the unique safety constraints of deploying clinical AI at scale. Ufonia’s conversational agent Dora has completed over 200,000 real patient calls across 20 UK hospitals—handling post-op and pre-op follow-ups—and the company is contracted to scale to one million patients within two years, with an active US expansion across six clinics in four states.
The core challenge Joselowitz addresses is that standard software de-risking playbooks break down entirely in healthcare. A/B testing is unethical and often illegal when patients are involved, calls cannot be rolled back once Dora has spoken, and benchmark scores from model vendors provide no legal protection after a clinical incident. In response, Ufonia built a simulation-first safety stack: a synthetic patient bot called PatBot generates thousands of diverse clinical dialogues, which are then evaluated by BevJudge, an LLM-as-judge system trained on documented clinical hazards developed with clinicians across multiple specialties.
BevJudge was validated against 240 expert-labeled examples from 10 clinicians across 10 clinical specialties. The best-performing model—Gemini 2.5 Pro—achieved an F1 score of 0.96 with near-perfect sensitivity on hazard detection, matching or exceeding expert clinician performance. A patient-public involvement study also found that simulated PatBot conversations were rated as more realistic than real doctor-patient dialogues in three out of four cases, validating the simulation approach as a credible pre-deployment safety harness.
📺 Source: AI Engineer · Published August 19, 2026
🏷️ Format: Keynote Launch
