Our Approach to Safety and Clinical Rigor

General-purpose AI will not heal the world.

It was built to answer questions and write code—not to guide a patient through months of recovery, not to recognize when someone is spiraling, not to know the difference between pain that means stop and pain that means push through. Each healthcare domain carries its own risks, its own standards, its own ways that AI can fail. Musculoskeletal care is not mental health, is not pregnancy, is not chronic disease management.

We believe healthcare AI must be built differently. Clinical expertise cannot be an afterthought. It must be in the foundation.

Clinical Co-creation as Methodology

We do not consult clinicians. We build with them.

Our AI research team works in direct partnership with clinical experts across every vertical we serve—physical therapists for Thrive, women's health practitioners for Bloom, licensed psychologists for Mind. They are co-creators from day one. They define what safe and effective care looks like, based on evidence, clinical guidelines and their experience. They design our evaluation frameworks. They validate that our AI behaves according to established professional standards and across diverse patient populations. If they are not at the table when systems are designed, we are building the wrong systems.

Domain-Specific Safety

Safety in healthcare is not one problem. It’s many.

A movement recommendation that ignores injury history can cause harm. A response to emotional distress that misses an escalating crisis can be dangerous. Guidance during pregnancy must account for gestational stage, medical history, and warning signs that demand immediate attention. General-purpose guardrails cannot make these distinctions. They were not designed to.

We build safety systems tailored to each clinical domain—grounded in the specific risk factors and standards of care that govern real practice. We take a comprehensive approach that factors in diversity in datasets, and thoroughly assesses for bias. There can be no shortcuts when lives are at stake.

Evaluation That Matches Reality

Most AI benchmarks are divorced from clinical reality. They test knowledge through static questions. They assess isolated responses. But healthcare unfolds over time.

Recovery from chronic pain involves progression across weeks. Mental health support requires maintaining therapeutic rapport across many conversations. Women's health spans conditions that evolve through life stages. If evaluation does not capture this complexity, we are measuring the wrong thing.

We develop evaluation frameworks that measure what matters: not whether an AI knows the right answer, but whether it can deliver appropriate care across the full arc of a patient's journey.

Continuous monitoring

Careful design and rigorous evaluation are not enough. Systems in production encounter the full complexity of real patients, real conversations, real life. They will fail in ways we did not anticipate.

We must watch. Continuously.

We build systems that analyze interactions at scale—clustering conversations, surfacing patterns, catching the failures that hide in the long tail. Not just tracking aggregate metrics, but understanding how our AI is being used and where it falls short.

This is how we find what we didn't know to look for.

Transparency and Scientific Contribution

We publish our research. We release our evaluation frameworks. We share our methodologies. This is not optional.

Advancing healthcare AI—particularly on questions of evaluation and safety—requires open scientific contribution. Transparency enables scrutiny. Scrutiny builds trust. Trust is the foundation of care. The principles that guide our work remain constant across every condition we treat: clinical expertise at the foundation, rigorous evaluation that matches real-world complexity, and safety systems designed for the domain they serve.