Evals-Driven Development for a Mental Health AI Coach — Akele Reed & Dave Revere, SonderMind

Testing a mental-health AI requires checking when it intervenes as well as what it says. In SonderMind's engineering talk, the care provider describes having clinicians annotate conversations with the expected response category and the point where a safety intervention should occur. Those annotations become repeatable tests for changes to prompts, models and guardrails.

The system separately checks incoming messages and outgoing responses. Its tests track missed interventions and inappropriate triggers, since either can undermine support. This makes clinical judgment part of the release process rather than a one-time review. The talk explains that engineering approach; it does not establish improved patient outcomes or prove the coach safe in every conversation.

“we were not going for more triggers here. We're going for more correct triggers.”

Akele Reed

“when you inappropriately guardrail on somebody, then that can often feel like a door slam to the face and make that person feel more isolated”

Akele Reed

“So, the point is the words aren't always the message. And that's the engineering problem I want to talk to you about.”

Dave Revere

“we all know that a simple eval gate does not make a system safe. A learning loop can.”

Dave Revere

“our system isn't deciding what correct is in a clinical edge case like this one. A licensed professional is.”

Dave Revere