Clinical Research AI Inside the Hospital Boundary

The operational problem

Research teams want retrieval over clinical text, but the governing constraint is that protected health information cannot leave the institution and every access must be attributable. That pushes the entire pipeline — embedding, retrieval, inference, and logging — inside the hospital boundary.

What Enfuse built

• On-prem inference and embedding services sized to the research workload
• PHI detection and redaction at ingest, with reversible tokens for authorized roles
• Study- and role-scoped retrieval with consent-status filtering
• End-to-end audit logging of prompts, retrieved documents, and outputs
• Evaluation harness for clinical extraction accuracy before each model change

Architecture

• Ingest — de-identification and indexing of clinical documents
• Retrieve — access-scoped RAG over the research corpus
• Reason — on-prem model inference with policy enforcement
• Record — immutable audit trail for compliance review

Hardware and software

• On-prem GPU servers
• Sovereign Runtime with DocuFlow retrieval service
• Hospital identity provider, SIEM, and existing EHR interfaces

Deployment environment

On-premises hospital datacenter. No external model API calls.

Results

• Retrieval-assisted review of clinical documents without data egress
• Attributable, auditable access for every request

Reference architecture. Describes the deployment pattern rather than measured outcomes at a named institution.