The problem

• Legacy clinical records system was a single VB.NET monolith on Windows Server 2012, written by three different vendors over a decade with no consistent data model.
• Six hospital sites maintained divergent local schemas; cross-site patient continuity required manual reconciliation.
• HL7 v2 messages from imaging (PACS), laboratory and pharmacy were processed by an undocumented BizTalk pipeline that nobody on staff understood.
• Compliance pressure: Ministry of Health audit findings on access control and audit trail completeness.

The solution

• Target-state architecture: .NET 10 service mesh with bounded contexts per clinical domain (records, imaging, lab, prescribing, scheduling). FHIR R4 as the canonical interchange model.
• HL7 bridge: Reverse-engineered the BizTalk pipeline into a documented, testable .NET worker handling HL7 v2 → FHIR translation, with a 6-month parallel-run safety net.
• AI-assisted triage: RAG layer over de-identified historical records assists triage nurses with similar-case retrieval. Generation kept advisory-only; no autonomous clinical decisions.
• Phased migration: Site-by-site cutover with strangler-fig pattern; each site validated for one month before next site went live.
• Compliance: Role-based access with break-glass audit, immutable audit log, encryption at rest and in transit. Ministry findings closed prior to next audit cycle.

Results

• All six sites migrated in 11 months with zero unplanned clinical downtime.
• Median chart-pull time across sites dropped from 14 seconds to under 2 seconds.
• Ministry of Health audit findings on access control and audit trail completeness fully closed.
• Triage staff reported the RAG-assisted similar-case panel as the single most-used new feature in post-launch survey.
• Project came in within the agreed 11-month window; phase-3 over-budget by 6% due to one scope addition (medication reconciliation) accepted mid-project.

Technology stack

.NET 10 (C#), FHIR R4, HL7 v2, PostgreSQL, Redis, Azure Service Bus, RAG over Postgres+pgvector, OpenAI / Claude / local Llama for triage advisory, Ministry of Health Israeli compliance framework.