Akele Reed is a principal AI engineer at SonderMind and an architect of Sonder, a clinically grounded mental-health companion that supports people between therapy appointments and connects them with human care when necessary. Reed designs the safeguards that determine when conversational AI can help, when it should provide additional resources, and when it must stop and direct someone toward immediate human support.
Reed has more than eight years of engineering experience and studied machine learning at Georgia Tech, with coursework in deep learning, computer vision, robotics, and AI ethics. At SonderMind, Reed works with clinicians, engineers, and product teams to turn mental-health expertise into practical safety infrastructure.
Building safer mental-health AI
- Modular clinical guardrails: Independent input and output checks surround Sonder Core, evaluating incoming messages, generated responses, and conversational context. Keeping these protections separate allows the central system to evolve while its safeguards remain independently testable.
- Context-sensitive crisis intervention: Active danger requires a different response from past trauma or relationship difficulties. Reed’s architecture distinguishes situations requiring urgent escalation from those where surfacing resources or continuing supportive conversation is appropriate; excessive blocking can isolate someone seeking help.
- Independent LLM-as-a-judge checks: Separate model-based safety assessments resist conversational manipulation more effectively than instructions embedded solely in the primary model. Additional latency and cost are deliberate trade-offs for stronger safeguards in a sensitive setting, as Reed explained during a conference session on mental-health AI safety.
- Clinician-informed safety evaluation: Licensed specialists help calibrate intervention thresholds, accounting for both missed dangers and unnecessary refusals. SonderMind’s open-source guardrail evaluation datasets include 200 input scenarios and 100 output scenarios spanning crisis disclosures, inappropriate diagnoses, medication recommendations, and other potentially harmful responses.
Reed approaches mental-health AI as a clinical safety problem: preserve access to support, recognize genuine risk, and return responsibility to qualified humans when the situation demands it.
