Data Fabric
Claims, FHIR/EHR, labs, pharmacy, notes, scheduling and provider operations converge into a longitudinal member record.
HEDIS, Star Ratings and CMS reporting
From Healthcare Analytics to Real-Time Quality Control
Nartis is a healthcare quality platform I designed and prototyped. It treats quality measurement as a control loop instead of a monthly report: clinical events are evaluated against measure policy as they arrive, open gaps are ranked and routed for action, and every decision keeps its evidence trail.
Event-drivenMeasure evaluation
Closed-loopGap closure
Audit-readyEvidence lineage
Outcome-ledStar Ratings impact
The problem
Quality teams find out too late.
Most quality programs run on batch claims, overnight ETL and dashboards that refresh weekly. By the time a gap shows up, the visit that could have closed it is over. Nartis is designed to evaluate each clinical event against measure policy when it arrives, rank the open gaps by what closing them is worth, route the right action to a person, and record why each decision was made.
Reference architecture
Legacy analytics vs Nartis
Legacy analytics
System of insight, retrospectiveLayer
Nartis
System of control, real timeClaims-centric data
Claims first, with EHR, lab and pharmacy feeds added later.
Batch feeds, daily to weeklyData
Multi-source events
FHIR/EHR, claims, pharmacy fills, labs, ADT and scheduling.
Streaming plus batchBatch ETL and warehouse
Pipelines clean and normalize, then load the warehouse.
Periodic processingProcessing
Event processing
Each event updates member state and re-checks measure eligibility.
Seconds to minutesMeasure logic in SQL
Rules embedded in pipelines. Hard to test, slow to change.
Static logicPolicy
Policy-as-code
HEDIS and Stars logic as versioned, tested rule packs.
Diffable, reviewable, auditableDescriptive analytics
Dashboards, KPIs and cohort reports.
What happened?Analytics
Predictive and prescriptive
Gap ranking, closure likelihood and next best action.
What should we do next?Static Star reporting
Point-in-time ratings with limited forecasting.
Backward lookingOutcomes
Star Ratings simulation
Distance to cut points and the bonus math behind them.
Forward lookingHuman-driven workflows
Care managers work lists outside the platform.
Manual hand-offsAction
Orchestrated actions
Provider alerts, outreach and scheduling, with a person approving anything clinical.
Closed loopTraditional governance
Role-based access and report-level audit logs.
User-focusedGovernance
Identity-aware governance
User, agent, purpose and data class checked on every request.
Purpose-bound, fully loggedDashboards and reports
Static views for executives and analysts.
VisibilityExperience
Role-based work queues
Ranked gaps for quality teams, action lists for providers.
ActionableLegacy: open loop
- Data
- ETL
- Warehouse
- Dashboard
- Manual action
Understands what happened.
Nartis: closed loop
- Event
- Policy
- Gap
- Action
- Outcome
- Feedback
Changes what happens next.
The shift: from retrospective reporting to an event-driven control loop with an evidence trail.
Operating model
Six layers of control
Processing Layer
Streaming and micro-batch pipelines keep member state and measure eligibility current.
Policy Layer
HEDIS, Stars and reporting logic live outside the pipelines as versioned, testable policy-as-code.
Intelligence Layer
Models estimate closure likelihood and value. A language model explains gaps and drafts outreach for review. It never decides measure status.
Action Layer
Provider alerts, member outreach, scheduling and care workflows close the loop.
Governance Layer
Identity, purpose, policy, data minimization and audit lineage control every action.
Positioning
System of insight vs system of control
Capability
Legacy analytics
Nartis
Core model
System of insight
System of control
Timing
Batch, retrospective
Event-driven, near real time
Measure logic
Embedded rules and ETL
Policy-as-code with tests
Stars view
Reporting and analysis
Cut-point simulation and intervention planning
Action layer
Depends on manual workflow
Orchestrated, with human approval for clinical steps
Governance
Traditional access control
User, agent, purpose and data-class controls
Trust layer
“The model never sees data unless both the user and the agent are authorized for that purpose, risk context, and data class.”
The platform treats AI as a governed actor, not a feature. Every request carries the identity chain across user, agent, data, policy, decision, action and audit event. In the prototype, the one live-model call sends synthetic fields only, with no names.
Architecture brief
Read the architecture brief
A short PDF covering the problem, the reference architecture, the six-layer operating model, the target AWS design, where generative AI fits, and the governance model.