Evals Driven-Development: Engineering a Mental Health AI Coach Ethically & Safely

Official Schedule Context

In the world of AI Mental Health, vibes can be dangerous with real consequences. Building

Sondermind’s Mental Health AI Coach required us to invent a new playbook for Eval-Driven Development

in order to balance effectiveness and safety. This session is for the builders who want to see how

to handle the most difficult edge cases in the agentic world. We’ll show how we’ve built a Clinical

Feedback Loop that turns human therapist insights into machine-readable evaluations in a production

system with thousands of conversations. We’ll dive into: - The Ethics Engine: Building and

calibrating modular guardrails that can be updated as clinical guidelines evolve. - Agentic

Oversight: Why we moved from single-prompt agents to a closed-loop Supervisor/Executor/Evaluator

pattern to ensure clinical adherence. - Human Oversight: How we monitor Sonder to ensure that we can

improve safety and quality with clinical feedback.

No official session recording transcript was found by exact title match on the AI Engineer YouTube channel during this run.

Pending transcript synthesis when an official recording or confirmed matching video is available.