Essays on clinician-governed AI
The Flywheel, Explained
Users generate data, data trains the model, the model serves providers, providers generate data. Ours is already turning.
Every AI company claims a data flywheel. Almost none has one that touches a real patient. A flywheel is easy to draw and hard to start: it requires live users, consented data, training infrastructure, and a product good enough that the improved model brings in more users. Here is ours, step by step, already turning.
Step one: the consumer layer. MPowerMe.app, live, with more than 3,100 real users across its apps: clinically grounded mood tracking with gratitude journaling, AI-enhanced sleep tracking, daily check-ins, PHQ-9 and GAD-7 instruments. Every interaction is real-world behavioral data, collected with consent under institutional review board governance.
Step two: purpose-built collection. That data flows through TherapyHub, designed from the start as training infrastructure, not scraped afterward from a product built for something else. Structure at the point of collection is what makes the data trainable.
Step three: the model improves. Reinforcement-learning training on real clinical data, under clinical supervision, in five languages. Each turn of the wheel makes Antiphon sharper at the tasks from our fourth essay: notes, intake, triage, planning.
Step four: providers adopt, and the wheel spins faster. The improved model ships as a HIPAA-compliant API. More providers build on it. More providers mean more data inflow, consented and governed, which means better training, which means a better model, which brings more providers. That is the flywheel, and it compounds.
That last distinction is the moat. Engagement data tells you what kept someone scrolling. Clinical data with provenance tells you who improved, under what supervision, in what context, measured how. One of those trains a better recommender. The other trains a better clinician’s assistant. There is no substitute for it, and there is no shortcut to it. You have to earn it one consented user at a time, which is what we have been doing, bootstrapped, with zero external capital.
The wheel is already turning. The capital we are raising now does not start it. It spins it faster: API rollout, provider adoption, international expansion across the five languages already live, university licensing for counselor training programs, and federal and grant contracts through our SAM.gov registered entity.
Governance: IRB IORG0012490, FWA00035466. Federal validation: CMS Health Tech Ecosystem Conversational AI Pledgee, listed at CMS.gov.