Regulated data
Where clinical or otherwise regulated data is in scope, the ingest path, encryption keys, retention and audit logging are agreed in writing before any build starts. Identifiers not needed for analytics are dropped at the boundary.
Reference architecture · illustrative
What the layer
underneath looks like.
Most organizations run on five to ten disconnected systems plus the spreadsheets filling the gaps. This page walks the pattern we build — land, conform, model, govern — using a healthcare example, where the requirements are hardest.
Example build
Invented values
Not a live environment
Read this as a worked example, not a product screenshot.
Every source, mapping, semantic entry, pipeline run and timestamp on this page is invented to show the shape of the build. It is not a deployed customer system, not connected to any live source, and not a statement about any customer's data. Platform, tooling and compliance posture are agreed per engagement rather than guaranteed here.
Architecture
Land it, conform it, define it once.
Source systems land as raw, versioned files or API pulls. A conformance step reshapes them into one canonical grain. The semantic layer defines each metric once, and everything above it — dashboards, exports, the assistant — reads that definition.
1 · Source systems
Clinical record systemFHIR R4 / HL7v2 exports
Practice management & billingScheduled extracts
HR information systemHeadcount, tenure, overtime
Payer portalsRemittance and quality files
General ledgerRevenue, expense, budget
SpreadsheetsThe parts no system owns
2 · Landing & conformance
Raw landing zoneImmutable, versioned, replayable
Managed clinical storeFHIR-native where clinical data is in scope
Conformed warehouseCanonical grain: entity × month × measure
Transformation with testsSchema, range and reconciliation checks
Scheduled orchestrationDaily, weekly and monthly runs
3 · Governed layer
Semantic catalogOne formula, unit and direction per metric
LineageAny number back to its source rows
Analytics surfacesDashboards, exports, APIs
Grounded assistantReads governed aggregates, never raw records
Platform choices
Chosen per engagement, not by default.
For regulated healthcare work we typically build on Google Cloud, because the services we need are covered by a Business Associate Agreement and the clinical store speaks FHIR natively. A smaller operator with no regulated data usually needs a far simpler stack, and we say so.
Tenant isolation
Each customer gets its own project or dataset, service identity, keys and audit trail, so cross-tenant movement is not a matter of configuration discipline.
Analytic grain
Operational analytics run on aggregates — entity × period × measure. That keeps the reporting layer small, fast, and free of record-level detail it has no reason to hold.
Assistants
Any assistant reads the governed aggregates and metric definitions, not raw source rows. The public demos on this site go further: they run locally with no language model at all.
The example build
Sources, mappings, metrics, runs.
The four views below show the artefacts a build actually produces. All values are invented.
Next step
See it applied, or map your own.
The industry demos show this same source-to-semantic structure running on synthetic data for sixteen industries. The assessment turns your answers into a sequenced recommendation in your browser.