Clinical AI · EPIC TECHNOLOGY
AI that can reason without becoming the source of truth.
I architected and built a clinician-facing AI platform for Ho Chi Minh City Traditional Medicine Hospital. The interesting part is not the chat box — it is keeping reasoning, evidence, clinical authority, state, failure behavior, and deployment constraints from collapsing into one opaque model call.
The engineering problem
A useful answer is not automatically a safe system.
The product has to understand natural language and synthesize useful responses, while application-owned facts, source provenance, permissions, conversation state, and clinical constraints remain independently inspectable.
I designed the system so AI reasoning can be flexible where it helps, but it cannot silently redefine authoritative knowledge or erase uncertainty when a dependency fails.
Public system cut
Four layers. Clear authority.
This is intentionally not the deployment diagram. It shows what each layer is responsible for, without publishing private corpus structure, prompts, provider configuration, infrastructure topology, or operational controls.
Clinical workflow
Questions, medication and herbal workflows, prescription review, follow-up context.
AI reasoning
Language understanding and synthesis operate inside bounded application contracts.
Evidence & authority
Governed knowledge, citations, deterministic facts, permissions, and durable state stay explicit.
Delivery boundary
Validation, failure behavior, deployment constraints, and human use are treated as engineering concerns.
What I owned
From ambiguous requirement to acceptance boundary.
Defined where language-model reasoning ends and application/source authority begins.
Built the clinician-facing workflow, APIs, durable state, retrieval/evidence path, permissions, and streaming behavior.
Kept dependency and provider failures explicit rather than converting them into plausible-looking clinical output.
Designed around on-premise constraints, testability, observability, provider portability, and clinician-facing acceptance.
Deliberately withheld
The public case study stops before the sensitive layer.
Private clinical corpus details, prompt/system instructions, internal schemas, credentials, deployment topology, machine identities, release artifacts, runbooks, clinician data, and exact operational controls are not published. The goal is to demonstrate engineering judgment — not ship a map of the hospital system.
The model may own synthesis. It should not silently own truth, permission, or evidence.
That boundary is the reusable part of this project: it applies to any high-consequence AI product where a fluent answer is weaker evidence than the system state behind it.
Explore more