Akele Reed describes SonderMind's mental-health AI coach as support before or between therapy sessions, with human care available when appropriate. Separate input and output guardrailsA guardrail is a technical or procedural control that limits unsafe, disallowed, deceptive, or otherwise unwanted AI behavior. surround the core conversational system, allowing modular testing at additional cost and latency. The talk includes discussion of self-harm, suicide and domestic violence.
Akele Reed distinguishes active danger from past distress and ordinary relationship difficulties. The engineering aim is context-sensitive intervention rather than more frequent refusals. Dave Revere explains how clinicians annotate ambiguous conversationsHuman-in-the-loop AI keeps a person involved in reviewing, correcting, approving, or guiding an AI system's work. with the expected intervention, risk category and turn index, turning expert judgment into typed regression evaluationsAI regression testing reruns preserved cases to detect whether a model or agent update has broken behavior that previously worked..
Dave Revere emphasizes false positives, false negatives and intervention timingAn AI evaluation metric is a defined measurement used to quantify a particular aspect of an AI system's performance., rather than optimizing an aggregate benchmark at the expense of users. The speakers announce 200 input and 100 output evaluation scenariosAn evaluation set is a collection of examples kept for measuring an AI system rather than training it. as a starting point, not a replacement for product-specific clinical review. The presentation describes a safety-engineering process; it does not independently establish clinical effectiveness or guarantee safe outcomes.
Watch on YouTube




