How SonderMind built safe AI coach: modular guardrails, clinical evals

SonderMind built Sonder, a mental health AI coach, using input/output guardrails with separate LLM judges. Modularity enables iteration on the core without compromising safety, a deliberate trade-off in latency and cost. The system is designed to avoid over-calibration—triggering correctly, not more often—by using clinical experts to define and annotate edge cases.

Each user conversation is traced; clinicians annotate failures, which become typed evals that gate releases. This human-in-the-loop process ensures that safety decisions are grounded in real patient data, not speculative prompts. The authors caution against pursuing perfect benchmarks, which can drift focus from human needs.

They open-sourced 200 input and 100 output guardrail scenarios to create a shared baseline, acknowledging that teams building similar systems shouldn't start from zero. The core insight: safety is not a static gate but a continuous learning loop owned by domain experts.

For engineering leaders, this demonstrates that AI safety in high-stakes domains requires organizational investment in expert reviewer pipelines, separate guardrail infrastructure, and evaluation frameworks that prioritize the right kind of calibration over raw sensitivity.