The challenge
Clinicians needed synthesis across records that sat in separate systems, but the regulatory floor made most off-the-shelf approaches unusable. Patient data could not leave controlled infrastructure and every inference had to be attributable.
Accuracy expectations were asymmetric. A missing suggestion is an inconvenience; a confidently wrong one is a patient safety event. The system had to be built to abstain rather than guess.
Approach
01
Compliance in the architecture, not the contract
Data residency, encryption, audit logging and access scoping were design constraints from the first diagram. Nothing was retrofitted after the build.
02
Hybrid retrieval over structured and unstructured records
Structured labs and vitals are queried directly; narrative notes and reports go through vector retrieval. The agent reconciles both rather than flattening everything into text.
03
Explicit abstention path
Below a confidence threshold the system returns what it found and states that it cannot synthesise a conclusion, rather than producing a lower-quality answer. This was specified before the build, not tuned in afterwards.
04
Clinician-graded evaluation
Accuracy was measured by clinical reviewers against real anonymised cases, not by automated similarity metrics.
Architecture
Ingest
HL7 and document feeds land in controlled storage inside the network's own cloud tenancy.
Protect
PII handling, encryption at rest and in transit, full audit trail on every read and every generated output.
Retrieve
Weaviate for narrative records with patient-scoped filters; direct queries for structured clinical data.
Reason
Claude Opus 5 on Vertex AI performs synthesis with mandatory source attribution per statement.
Review
Every output surfaces its sources so the clinician can verify before acting. The system supports, it does not decide.
Results
91%Factual accuracy in production
27%Reduction in intake time
48%Less manual review
85%Fewer false alerts
HIPAACompliant by design
What I took from it
Building the abstention path first changed the whole project. Once the system could say 'I cannot conclude this', clinicians stopped treating it as a black box and started treating it as a colleague that flags uncertainty.
HIPAA-compliant patient monitoring AI reduced false alerts by 85% across our hospital network. They understood clinical constraints, regulatory requirements, and technology simultaneously.
Stack
Models
Cloud
Data
Compliance