Revenue analytics platform

Context

• Leadership asks the same question three ways and gets three answers, each defensible, none reconcilable.
• Every team has built its own extract. The finance number, the sales number and the board number diverge, and reconciling them is a monthly manual ritual.
• Nobody disputes the data is wrong. What is disputed is whose version is wrong.

What was actually wrong

• The failure is definitional, not technical. Terms like net new, active customer, qualified, churned and booked are each used with several meanings, and no artifact records which meaning a given report used.
• Dashboards are built directly on source extracts, so business logic lives inside a dozen visualisation layers where it cannot be tested, reviewed or reused.
• Because the logic is invisible, a change to it looks like a change in the business.

Approach

1. 01Start from the decision, not the data: inventory the decisions leadership actually makes on a cadence, and work backwards to the smallest set of metrics that move those decisions.
2. 02Write the definitions down as a metric contract - grain, filters, inclusion rules, effective dates, owner - and get them ratified by the function that owns the number rather than by the analytics team.
3. 03Model in layers: raw landing kept immutable, a conformed core with slowly-changing dimensions handled explicitly, then presentation models shaped to the questions.
4. 04Serve every surface from the same semantic layer, so a dashboard, an export and an API answer identically by construction.
5. 05Close the loop with a governance cadence: definition changes are reviewed, versioned and announced, because a silent definition change is indistinguishable from a business event.

Architecture

• Cloud data warehouse with partitioned, clustered fact tables sized to the real query patterns
• Transformation as version-controlled, tested models with lineage from source to serving
• Slowly-changing dimensions handled explicitly, so historical questions get historical answers
• A semantic layer as the single definition of every metric
• Containerised web application for the serving surface, deployed through CI with identity-aware access

What this demonstrates

• Turning an argument about numbers into a governed system with owners
• Warehouse modelling that survives contact with real history rather than only the current snapshot
• Designing for reconciliation up front, so the finance view and the board view cannot silently drift

Related