Built like a medical platform.
Operated like software.
The differentiator isn't cheap operations. It's clinical-grade safety architecture on a multi-tenant platform. The same codebase serves D2C members, employer groups, and white-label partners. Click any layer to explore.
Operating leverage
Platform and hosting run for about $4K a month and stay under $5K at 10,000 members. AI handles first-line member support and clinical admin; people handle escalations. 22 named roles by the end of year one, each hired only where AI cannot do the work.
Clinical-grade safety architecture
This is why an employer or health system can buy Halftime rather than merely link to it, and why a storefront competitor can't claim it. The platform is built multi-tenant from day one, with deterministic safety controls running ahead of the model.
tenant_id on every row, Postgres row-level security on all clinical tables, ltree ancestry scoping for network partners managing sub-tenants. 7 tenant types: D2C, venue, corporate, affiliate, PEO, white-label, network. Envelope: 1,000 tenants × 100K users each.
partner.halftime.health, tenant-scoped dashboards, member lists, referral tracking, employer-group eligibility management. QA-audited 20 June against a real tenant.
Crisis and emergency detection runs before the LLM is invoked, routing to human escalation rather than to a model. Almost nobody in consumer health has this.
Structured marker parsing → reference-range zone classification → Bedrock protocol recommendation → physician review, with analysis-guardrails and an approved-protocol check before fulfillment can fire.
Every state change writes an OutboxEvent in the same transaction; BullMQ workers drain it; retry ceiling with a dead-letter triage UI. Consistent across clinical, billing, and ops.
Ordering, lab, and fulfillment integration that re-validates eligibility and protocol state immediately before any shipment fires, a safety pattern, not a batch sync.
Including ElevenLabs, with a "never log PHI or transcripts" contract enforced in code. Drata for compliance automation across the full stack.
The behavioral intelligence layer
Every patient interaction will generate signal. Over 12–18 months this is designed to accumulate into a proprietary dataset no competitor can replicate without building an equivalent patient base from scratch. The systems below are designed and wired, not yet running on live patient data.
High expansion revenue potential. Selects higher tiers immediately. Wants quantified progress data.
Price-conscious but pays for results. Evaluates on tangible outcomes within 60 days.
Browses extensively. High lifetime value through protocol expansion. Social sharer.
Arrives cautious. Starts with lowest tier. Critical first 30 days. Very loyal after month 3.
Generates 3–5x avg referrals. Becomes content contributor. High brand value beyond subscription.
Designed inputs: refill timing, engagement velocity, support ticket sentiment, content consumption drop-off, outcome self-report trends. Designed output: a risk score updated weekly, with scores above threshold triggering automated intervention 30 days before churn intent.
Patients like you who've been on RECOVER for 3 months typically add DEEP REST for sleep optimization. Recommendations are designed to feel like clinical guidance backed by aggregate outcome data, not sales pitches.
For each protocol, the engine is designed to build effectiveness profiles by behavioral segment: avg satisfaction score, time-to-first-improvement, retention at 3/6/12 months, most common protocol additions.
The engine is designed to learn each patient's preferred channel, frequency, timing, and content type. The Optimizer gets data-rich weekly updates. The Pragmatist gets concise monthly check-ins.
Each patient tagged within 30 days based on behavioral signals. Communication cadence and content personalized.
Enough cohort data to predict churn 30 days in advance with meaningful accuracy. Intervention engine activates.
Dataset includes behavioral profiles across thousands of patients and 15 protocol types. No competitor can replicate without their own patient base.
Predictive models calibrated with real outcome data. A new entrant would need 24 months of operation to approach equivalent model quality.
Every manual process is a workflow waiting to be automated.
AI structures notes → physician alert → protocol draft generated
Prescription sent to WellSync → shipping triggered → patient notified
AI interprets → dashboard updated → physician review flag raised if out-of-range