Health Data Agent

Health Data Agent is a working healthcare AI MVP and governance case study. It demonstrates HL7 to FHIR processing, patient-data chat, clinical summaries, and useful AI workflows for healthcare data.

This is not only a data-protection problem. When AI touches sensitive workflows, one hallucination, missed detail, or flawed summary can influence decisions that affect real people: their care, finances, rights, safety, or reputation. If that mistake causes harm, the organization is not just facing a technical failure. It may be facing lawsuits, regulatory scrutiny, and legal exposure that can threaten the business itself. AI governance is what makes that failure defensible with an audit trail.

Warden 1.0 demonstrates the first version of the safety story: PHI tokenization, basic prompt guardrails, policy checks around tool calls, and PHI-conscious audit events. Those controls are real, but the demo intentionally shows that they are not a complete governance system.

• Identity gap: a system cannot assign responsibility if it does not know who is using it.
• Scope gap: clinicians may need PHI, but access still needs role, patient, and purpose boundaries.
• Provenance gap: clinical ingestion needs trusted sources, not just valid-looking HL7.
• Token gap: internal safety markers must not become predictable decoder rings.
• Audit gap: governance logs must help answer who accessed or changed what, why, and under which policy.

Real AI governance starts by making failure modes visible. Warden 1.0 is useful because it looks like the kind of safety layer many AI products advertise: privacy language, guardrail language, audit language, and an impressive demo. The lesson is that those claims need enforceable controls behind them.

This MVP is incomplete by design so the next version has a clear target. The demo shows the app, the risk, and the fix path in one place.

The core risk is not that clinicians can see PHI. Clinicians often need PHI. The liability question is whether the system can prove the access was authorized, scoped, necessary, sourced from a trusted workflow, and recorded in a useful audit trail.

Health Data Agent uses the current MVP to make that question concrete. When an AI system connects natural language to clinical data, ingestion workflows, and summaries, governance becomes the control plane around every capability.

Warden 2.0 is the hardened build and architecture for turning this from a compelling MVP into a production-minded governance system. It treats AI safety as an access-control and accountability problem around the model, not a prompt-only problem inside the model.

Hi, I'm Bradly Cheng, a software developer and AI engineer focused on healthcare AI governance, interoperability, security, and practical clinical AI systems.

I built Health Data Agent to show that building the AI workflow is only half the work. The more serious challenge is identifying where the first safety story fails, then designing the controls that make the system attributable, scoped, auditable, and production-minded.

Warden 1.0 is the visible demo layer. Warden 2.0 is the direction I care most about: governance that can stand up to real questions about liability, trust, PHI minimization, and clinical data integrity.

Healthcare is the test case because the stakes are obvious, but the same governance problem applies to any AI system working with sensitive data: finance, identity, legal records, private communications, internal company data, or any workflow where the output can affect real people.