Hospital Clinical Intelligence PlatformCapability → responsibility → implementation technology
HOSPITAL APPLICATION BOUNDARY MODEL-VISIBLE CLINICAL INTELLIGENCE Clinicianquestion · files · follow-up Clinician web shellinteraction · streaming UINext.js · React · TypeScript Authenticated APIconversation boundaryPython · FastAPI · SSE/WS Hospital platform shellACL · audit · durable statePostgreSQL · Python kernel Clinical intelligencesingle model-visible rootPython · LLM tool orchestrationOWNED BOUNDARY Evidence Workspacecanonical observations · provenancePostgreSQL Clinical fact toolsdeterministic approved factsPython · typed services Evidence researchbounded approved corpusPython · hybrid retrieval Vector / hybrid searchderived · rebuildableQdrant · pgvector Main clinical agentsingle semantic ownerLLM · Python bounded loopBOUNDED LOOP Evidence verifier + guardsconditional exact-support checkPython · deterministic guards Answer + citationspersisted · auditablePostgreSQL · provenance Async & deployment runtimebackground work · protected/on-prem service deliveryCelery · RabbitMQ · Docker · Linux · Nginx
Reading order:capability→ responsibility→ implementation technology
Single semantic owner
  • LLM reasoning is bounded inside a Python-owned tool loop; it does not own clinical truth.
  • FastAPI/PostgreSQL application contracts retain permissions, state and provenance.
Evidence & authority
  • PostgreSQL is canonical durable truth; Qdrant/pgvector search is derived and rebuildable.
  • Deterministic clinical services remain non-redecidable by the model.
Runtime depth
  • Next.js/React provide the clinician surface; FastAPI owns the service boundary.
  • Celery/RabbitMQ handle asynchronous work; Docker/Linux/Nginx support constrained deployment.
Ho Chi Minh City Traditional Medicine Hospital · Clinical AIArchitecture and technology choices shown at capability level