Jared Joselowitz

You can't A/B test on patients. You can't unsend a phone call. The model card won't save you at the post-incident review. Most AI eng playbooks assume the opposite. Ship to 5%, watch the dashboard, roll back if it goes wrong. None of it survives regulated deployment, which is now coming for fintech, legal, and government too. So the engineering has to move: into hazard analysis, simulated populations, asymmetric evaluation, and audit trails treated as the deliverable. The trail is the product. I'll show you what changes when rollback isn't an option. How Ufonia ships Dora, an AI voice agent now making clinical follow-up calls on the NHS and across US health systems, using a hazard-driven simulation rig (MATRIX) and a prompt-optimisation flywheel that surface failures and conform the same base system to each clinical niche, all of it pinned to an audit trail. And the cheap version of all this, for any team whose users can't be the test population.