Dave Revere builds clinician-informed evaluation systems for mental-health AI, translating professional judgment into engineering safeguards for Sonder, SonderMind’s AI mental-health companion. At AI Engineer World’s Fair 2026, he was identified as a staff AI engineer at SonderMind, working on technology designed to complement therapy and direct people toward human care when necessary.
His approach turns difficult conversations into durable product improvements. Clinicians review interactions, specify the appropriate response, identify when an intervention should occur, and provide structured annotations that become automated regression tests. Changes to prompts, models, or guardrails can then be checked against accumulated clinical expertise before release.
- Clinical judgment embedded in release testing. Licensed professionals define the correct response to ambiguous situations; engineers convert their assessments into typed evaluations covering risk categories, expected interventions, and conversational timing.
- Context-sensitive mental-health guardrails. Indirect expressions of distress can evade keyword filters and generic moderation, while excessive intervention can interrupt constructive conversations. Revere emphasizes calibrating both false negatives and false positives so users receive continued support, additional resources, or escalation as appropriate.
- Human-centered safety benchmarks. His evaluations prioritize clinically meaningful failure modes over impressive aggregate scores, recognizing that mental-health conversations often contain ambiguity that standardized benchmarks alone cannot resolve.
- Open-source guardrail calibration datasets. Revere helped introduce SonderMind’s public evaluation repository, which contains 200 input scenarios and 100 output scenarios developed with clinical review. The materials offer other builders a starting point for testing conversational safeguards, without replacing professional oversight or product-specific evaluation.
In his AI Engineer World’s Fair presentation, Revere connects these practices to a clear operational principle: mental-health AI earns trust when licensed experts shape its safeguards, their judgments survive as repeatable tests, and human care remains available when software reaches its limits.
