Case study · CruCare Health
Four milestones in eight days, and one rescue
A supplement protocol that changes between two identical intakes is useless in a clinical setting. Making a language model reproducible, and making its safety rules unbreakable, was the job.
The problem
CruCare Health runs a supplement retail chain in Metro Detroit serving Medicare Advantage members who pay with an over-the-counter benefit card. A consultant runs an entire in-store visit by hand: intake, a recommended protocol, a physician review, payment across four different benefit terminals, and months of follow-up.
They wanted the visit run by software, with the protocol recommended by AI. In a health setting that raises two questions immediately. Will it give the same advice twice? And what stops it recommending something that will hurt someone?
Two answers we engineered
The AI gives the same answer twice
The model runs at temperature zero behind a content-hash cache, so an identical intake always produces an identical protocol. Doses are taken verbatim from the product catalogue rather than paraphrased by the model. The model is constrained to a per-visit allow-list, with a second independent filter that rejects any product outside it. Recommendations are grounded in the client's own dosing spreadsheets, embedded and retrieved per consultation.
Safety rules are code, not prompt text
Kidney disease caps vitamin D3 at 1,000 IU. Blood thinners flag fish oil, CoQ10 and vitamin E. Pregnancy gates the entire protocol.
These are enforced in the system, not requested of the model. A prompt can be talked out of a rule. A conditional cannot. Any safety constraint that matters belongs outside the model.
What shipped, and how fast
The scope of work allowed 14 days. Four milestones landed in eight.
| Milestone | What it added |
|---|---|
| Two | Five-step consultation wizard, 28-field intake with auto-save, benefit-card detection, physician review with video escalation, four-terminal split payment |
| Three | Client records with a progress timeline, a real-time walk-in queue, and a separate consultation engine for pets that doses by weight |
| Four | Public booking widget, KPI dashboard, inventory synced with the point-of-sale system, automated follow-up sequence |
| Five | Retention engine across a 30-day lifecycle, acquisition analytics, and a permission matrix of 29 features across five roles |
Then a line-by-line audit against the client's own 4,196-line prototype specification, which found eight gaps and closed all of them.
Built with React and Node on PostgreSQL with pgvector, integrating payments and catalogue sync, SMS, HIPAA-compliant video, and email. HIPAA and TCPA consent gates throughout, a full audit log carrying user and IP, and no protected health information in logs.
The part we are proudest of
In June the model vendor retired the version every consultation depended on. The AI stopped working across the entire platform.
We restored every consultation, then shipped an automatic safeguard so that class of failure cannot repeat. The month's report recorded one critical outage restored and everything deployed to production.
Any vendor can look good while nothing breaks. What a client is really buying is what happens on the day something does. The engagement continued onto a monthly retainer after the fixed-price milestones ended.
Have a model that needs to behave?
We build AI systems that are checkable: grounded answers, reproducible output, safety rules in code, and tests that gate every deploy. Tell us what is breaking and we will tell you whether we can fix it.
Start a conversationOr email info@codezbit.io