World's Health Platform

A cardiovascular care platform where AI drafts and physicians decide. Lab results run through a deterministic interpretation engine, an LLM writes the physician brief from a pseudonymised packet, and nothing reaches a patient until a physician signs it. In development.

Built at: DevTechGuru
Scale: 14 bounded contexts and 27 user stories designed as one governed pipeline: intake, lab results, interpretation, AI brief, review, sign-off, release and follow-up.
Category: Healthcare AI

Bounded contexts: 14
Identifiers sent to the model: 0

Context

Cardiovascular care runs on lab results. A physician has to turn markers, bands and history into a plan the patient can act on, and the client wanted AI in that loop without letting it decide anything.

Challenge

Health data is special-category data, and a language model is the least predictable component in the system. The AI had to be useful to a physician without ever seeing who the patient is or deciding anything on its own.

Non-negotiables

The model never sees patient identifiers
No plan is released without a physician signature
Every read and change of health data is audited

Calls I made

01
Deterministic engine, AI for prose only
Thresholds, bands, composites, rules and ranking are versioned clinical content executed by plain code. The model only drafts the physician brief from the approved packet, so every clinical conclusion can be replayed and explained without it.
02
A pseudonymised packet behind a port
The drafting context swaps identifiers out through a pseudonym vault before anything is composed, and the model sits behind a single composer port. The default adapter is an offline template; the Anthropic adapter switches on only when a key is configured.
03
A blocking validator before the physician
Every draft passes an output validator that can stop it before a physician sees it. Only a physician-signed plan can be released, and only by an explicit action. There is no auto-release.
04
Encrypted, tenant-scoped, audited
Identifying and clinical values are AES-256-GCM encrypted at rest, every query is tenant-scoped, every change emits an audited domain event, and reads are recorded too.

Trade-offs accepted

Keeping the model out of every decision means the rules engine is written, reviewed and versioned by hand. That is slower than letting a model infer the rules, and it is the point.
All clinical content in the build is synthetic and labelled as such, so nothing here has been measured against real patient outcomes yet.

Outcome

In development. The architecture and the governed pipeline are designed, and the platform is being built slice by slice, each one leaving the full test suite green. Every clinical record in it so far is synthetic.

The safest place for a language model in a clinical system is the narrowest one. Give it prose to write, never a decision to make, and put a validator and a physician between it and the patient.
What I took away

Stack