How Vipra Software unified 47 disparate EMR systems, DICOM medical imaging, and unstructured physician notes into a single AI-native data fabric — enabling clinicians to query the entire patient universe in plain English and reducing diagnosis time by 40%.
A national healthcare network spanning 22 hospitals and 180+ outpatient clinics had grown through a decade of acquisitions. The result: 47 distinct Electronic Medical Record systems — Epic, Cerner, Meditech, Allscripts, and 43 smaller vendor deployments — each siloed, each holding fragments of a patient's true clinical story.
The consequences were measurable and dangerous. Physicians making critical decisions had access to partial records. A patient presenting at Hospital B with atypical symptoms might have diagnostic notes from Hospital A sitting in an unreachable system. Drug interaction checks ran only against the local pharmacy database — not the patient's full medication history across all facilities. Radiologists interpreted DICOM scans without the context of the patient's written clinical narrative.
The organisation's AI initiative stalled for the same reason all healthcare AI fails: you cannot train or run a diagnosis assistant against data that doesn't exist in one place. Structured labs, ICD-10 codes, and vitals were in one set of databases. Unstructured physician notes, radiology reports, discharge summaries, and 50M+ medical literature papers were in another world entirely — unsearchable, unqueryable, invisible to any analytics tool.
85% of clinically actionable data in healthcare is unstructured — physician notes, imaging reports, discharge summaries. Traditional analytics platforms can only see the 15% that's structured. The Clinical Intelligence Fabric was built to close this gap at scale.
47 EMR systems with no common patient identifier, schema, or interoperability standard. Manual reconciliation estimated at 4+ hours per patient transfer.
6M+ physician notes, 2M+ radiology PDFs, and 50M medical research papers inaccessible to any existing analytics or decision-support tool.
Every vector embedding, LLM inference call, and RAG retrieval operation must meet HIPAA technical safeguard requirements. Standard LLM APIs are non-compliant by default.
Clinical AI that generates plausible but incorrect diagnostic guidance is a patient safety liability. Standard Gemini outputs needed strict grounding against verified medical data.
The platform is built on three foundational layers: a unified data ingestion fabric (structured + unstructured), a Vertex AI–hosted Gemini Pro inference layer with RAG grounding, and a HIPAA-compliant conversational interface for clinical users.
The core innovation of the Clinical Intelligence Fabric is its ability to treat structured FHIR data and unstructured clinical text as first-class citizens in the same query pipeline. Here's how a single clinical query traverses all layers:
The project was structured in four phases, each building on the last, with HIPAA compliance threading through every sprint rather than being bolted on at the end.
All clinical data — regardless of source system schema — is transformed to FHIR R4 resources before entering the platform. This single canonical model enables the RAG engine to query across all 47 source systems as if they were one database.
Base Gemini Pro was fine-tuned on 2M curated clinical Q&A pairs derived from the organisation's own de-identified case library plus MedQA, MedMCQA, and PubMedQA benchmarks. Critically, fine-tuning was performed within the HIPAA VPC boundary on Vertex AI — patient data never left the compliance perimeter.
Fine-tuning over RAG alone: the model needed to understand clinical reasoning patterns, not just retrieve facts. RAG provides grounded evidence; fine-tuning provides clinical judgment. Both are required for safe clinical AI.
Every Gemini response passes through a three-layer validation pipeline before reaching the clinician: (1) citation coverage check — every claim must be attributable to a retrieved source document; (2) drug interaction cross-check against the RxNorm database; (3) confidence scoring with automatic escalation to human review below a configurable threshold.
No shared patient ID. Built a probabilistic MPI using Soundex + Levenshtein name matching + DOB + ZIP code, resolving 1.2M duplicate records with 99.1% accuracy validated by clinical audit.
Standard LLM APIs log prompts. Deployed Gemini Pro on Vertex AI within VPC Service Controls with Customer-Managed Encryption Keys (CMEK), ensuring PHI never crosses the HIPAA boundary.
Naive ANN search over 680M vectors was too slow for clinical use. Built a two-stage retrieval: fast keyword pre-filter (Elasticsearch) narrows to 50K candidates before Matching Engine ANN achieves <400ms P95.
Physicians didn't trust black-box answers. Implemented full citation UI showing every source document used — physicians can click through to the original note, lab result, or research paper behind any AI statement.
Radiology PACS systems don't speak FHIR. Built a DICOM-to-FHIR bridge extracting study metadata, linking to patient FHIR records, and embedding radiologist impression text for vector search.
Clinicians must only see their facility's patients. Implemented BigQuery row-level security and Vertex AI Matching Engine namespace isolation ensuring zero cross-facility data leakage, validated by penetration test.
Don't build custom schemas for clinical data. FHIR R4 is the lingua franca of modern healthcare interoperability. Build to it from day one and every downstream tool — AI, BI, and interoperability APIs — benefits.
Compliance controls must be embedded in the data pipeline design, not added post-launch. Every VPC, every encryption key, every audit log should be designed in Sprint 1, not Sprint 20.
RAG alone doesn't give the model medical reasoning ability. Fine-tuning alone hallucinate facts not in training data. The combination — RAG for grounding, fine-tuning for domain reasoning — is what makes clinical AI safe.
Clinical AI that doesn't show its work won't be trusted or used. Every Gemini response should cite its sources with links to the original documents. Trust is earned through transparency, not accuracy scores alone.
Configure confidence thresholds below which AI responses are flagged for physician review. Clinical AI is a decision-support tool, not a decision-making tool. Design the system with this boundary explicit.
For corpora of 50M+ documents, ANN vector search alone at full scale is prohibitively slow. A pre-filter stage (keyword/BM25) that narrows the candidate set by 99% before semantic search is the production-proven pattern.
The Clinical Intelligence Fabric went live across 8 pilot hospitals in Week 30, with full network rollout completing in Week 38. Within the first 90 days, clinical outcomes data confirmed the projected impact across every target metric.
The drug interaction reduction was the most immediate and measurable impact. Pre-launch, the organisation averaged 18 adverse drug interaction events per quarter that were flagged in post-incident review as "preventable with complete medication history." In Q1 post-launch, this fell to 7 — a 61% reduction attributable directly to the fabric's ability to surface the patient's complete cross-facility medication history at point-of-care.
The $2M+ annual savings figure is conservative. It accounts for reduced readmission costs ($1.1M), clinical staff time recovered from manual record reconciliation ($620K), and reduced duplicate diagnostic tests ordered due to physicians not knowing prior test results existed in another system ($340K). The avoided cost of one serious adverse event that would have generated regulatory liability is not included in this figure.
The FDA is fast-tracking AI diagnostic tools under its Digital Health Center of Excellence framework. Organisations that deploy HIPAA-compliant, citation-grounded clinical AI today are positioned to file for FDA SaMD designation as the regulatory pathway matures — creating a defensible moat against competitors still operating on fragmented data.
The Clinical Intelligence Fabric architecture is replicable. Talk to the engineers who designed it.
Talk to an Engineer → More Engineering Projects