Clinical AI Gateway: a control plane for PHI-safe AI.

The problem

Healthcare organizations want AI assistants for billing, clinical workflows, the front office, and patient support. What stops them is never the model; it is everything around the model. Where does PHI go? Who reviewed the output? What did this tenant spend? What exactly happened when the auditor asks, months later? A model call is one line of code. The governance around it is the actual engineering problem.

The design

The gateway is a control plane between the assistant fleet and the model layer. Every request passes through the same governed pipeline in the same order: a verified JWT carries the tenant and scope, usage metering admits the request against per-tenant budgets, the redaction service strips PHI before anything reaches a model, governance policy decides whether a human reviews the output, and every state transition is published to Kafka and persisted to a dedicated audit read model, decoupled from the write path.

The default posture is autonomy with observability: hard budgets and human review are policies an operator turns on where risk justifies them, not fixed gates on every request. The model call itself sits behind a vendor-neutral provider seam, so the same governance plane works whichever model vendor the organization chooses, or switches to later.

The engineering

Java 21 and Spring Boot 3, because that is what healthcare enterprises run and trust. Thin controllers, a single orchestrating service that owns the governed lifecycle, PostgreSQL with Flyway migrations, Redis for metering counters, Kafka for the audit stream. The architecture ships with five volumes of documentation: the technical architecture, diagrams, architecture decision records, an enterprise standard, and model and redaction research. The decisions are written down, with the reasoning, before the auditor or the next engineer has to guess.

Why it matters to your project

This is the compliance-first pattern we bring to regulated work: data minimization before the model, audit trails as a first-class subsystem, policy as configuration rather than hard-coded gates, and a vendor seam that keeps you portable. If your customers handle PHI, PCI, or anything an auditor will one day ask about, the governance plane is the part of the system that decides whether AI is deployable at all. We design it first, and we have the reference implementation to prove the pattern works.

The architecture is open. The code, the diagrams, and the decision records are public: read the Clinical AI Gateway on GitHub.