Skip to content
Product we builtAI ImplementationProductAI

A Compliant Wellness Plan in 63 Seconds

True Health, a multi-tenant plan platform for wellness clinics we designed, built, and shipped · Apr–Aug 2026

  • Supabase
  • Custom
3m13s → 63.5splan generation
63.5splan generation, down from 3m13s
886automated tests, up from 480
~30%estimated lower cost per plan through caching
30KBAPI payload, down from 400KB

Why We Built It

True Health started as a single-operator wellness practice's intake problem and became a product. A single-operator wellness practice ran intake on phone calls and sticky notes and wanted plans that didn't "read like a chatbot wrote them." We built the platform, then pivoted it to a multi-tenant product for clinics after the first working version.

What It Does

  • Staff fill a 15-field intake and upload lab work. The platform produces one coherent, compliant client plan in about a minute, as a branded PDF.
  • Intake is matched to a methodology by the model with a confidence score; a deterministic ranker picks the practitioners and staff confirm every match before a plan runs.
  • Stage 1 writes an educational knowledge brief per practitioner methodology with no client data, and caches it. Stage 2 matches the brief to the client's context. Stage 3 synthesizes one plan and resolves conflicts decisively.
  • A refusal detector, a retry ladder that generalizes sensitive inputs, and a deterministic fallback guarantee a usable plan every time.
  • Lab PDFs are transcribed, never interpreted: values as strings, only the lab's own flags, enforced by schema.
  • A compliance layer combines prompt rules with a deterministic scrubber for banned words, named protocols, and practitioner-name leaks, tested against an adversarial golden set that never calls the model.
  • Multi-tenant from day one: row-level security, super-admin impersonation, audit log, quotas, and per-tenant branding.

Results

  • Plan generation from 3 minutes 13 seconds down to 63.5 seconds.
  • Automated tests grew from 480 to 886 in one session.
  • An estimated 30 percent lower cost per plan through caching.
  • Deployed to production, with a live demo to clinic staff in August 2026.
Where AI does the work

Synthesizing methodology into a brief, matching it to a client, writing the plan, transcribing labs, proposing the methodology match, with a confidence score, that a person confirms.

Where rules do the work

Compliance scrubbing, contraindication checks, refusal handling, the practitioner ranker and the confirm click on every match, tenant isolation, quotas, and everything in the golden set. The model never gets the last word on a compliance question.

The mechanisms behind it

AI Inside the WorkflowShippedAI

Rules get the last word on every model-drafted plan

Rules rewrite banned words, named protocols, and practitioner names in every plan before it is stored. Ten adversarial fixtures prove it with no model call.

  • Supabase
  • Custom
Read the use case →
AI Inside the WorkflowShippedAI

Lab reports to named markers, transcribed and never interpreted

A model reads the lab report as data entry, not analysis. Values stay strings, only the lab's own flags count, and the schema has nowhere to put a diagnosis.

  • Supabase
  • Custom
Read the use case →
AI Inside the WorkflowShippedAI

The model proposes, a rule ranks, a person confirms

The model proposes up to three scored methodology matches. A ranker with no model calls picks practitioners. Staff confirm every match before a plan runs.

  • Supabase
  • Custom
Read the use case →
AI Inside the WorkflowShippedAI

Three stages to one plan, with the brief cached once

Stage 1 writes a methodology brief with no client data, cached by content hash. Stages 2 and 3 match and synthesize in a background job that streams a preview.

  • Supabase
  • Custom
Read the use case →

Facing a similar challenge?

Let's talk about what's going on with your systems. 30 minutes. No pitch deck. We'll tell you if AI is the wrong tool.

Book a Discovery Call

We'll respond within 24 hours.