Descriptions:
Akele Reed and Dave Revere from SonderMind present the engineering architecture behind Sonder, a clinically grounded AI mental health coach deployed to serve users between therapy sessions or before they are ready to seek professional care. With SonderMind having served over one million people and citing American Psychological Association data showing 77% of psychologists report patients using AI for mental health support, the talk addresses a real and growing gap in the market.
The core architecture sandwiches Sonder’s central LLM between independent input and output guardrails implemented as separate LLM-as-a-judge calls. This modularity — despite adding latency and cost — makes the system significantly harder to jailbreak through sustained conversational pressure. Input guardrails screen messages before the core responds; output guardrails review the full conversation holistically and can intervene when clinical safety appears at risk. Every architectural decision, the speakers emphasize, was made with safety as the primary objective over performance or cost.
Dave Revere then focuses on the evaluation methodology that makes these guardrails trustworthy in practice. The key challenge is detecting indirect or coded distress signals that word-matching and broad moderation APIs miss entirely — illustrated by an example drawn directly from a clinician’s real patient experience. The team builds its evaluation loop in partnership with practicing clinicians, who provide calibrated examples of genuine risk signals, enabling the system to learn to detect the meaning beneath the words rather than surface-level keywords.
📺 Source: AI Engineer · Published July 25, 2026
🏷️ Format: Deep Dive
