Who it's for
Companies whose numbers live in spreadsheets and app databases, where each team calculates revenue its own way. Also companies that want AI tools to answer data questions correctly, because those tools are only as good as the data models and definitions underneath.
What we deliver
The right architecture for the size of the problem
Warehouse or lakehouse, ELT or ETL, batch or streaming. On Databricks, Snowflake, BigQuery, Redshift or Fabric, or on Postgres and DuckDB for small teams.
Data models people and AI can read
- Dimensional models built with Kimball's four steps: business process, grain, dimensions and facts.
- A declared grain for every fact table.
- History kept with Type 2 slowly changing dimensions where it matters.
- Conformed dimensions shared across teams.
- Clear table and column descriptions.
Pipelines that catch problems before users do
- Raw, staging and mart layers (bronze, silver and gold) built with dbt.
- Loads that pick up only new files.
- Tests on keys, relationships and accepted values, plus freshness checks.
- Model contracts on shared tables.
- Documentation written into the catalog.
Orchestration
The run order, schedules, retries and an owner who gets the alert, plus a plan for bad batches: fix, filter or quarantine.
Governance
Access by role, column masks and row filters for personal data and a separate, restricted identity for each AI agent.
Cost control
Right-sized compute, incremental loads and sampling before running AI over large tables.
How we measure it
- Tests passing and freshness targets met.
- Totals that reconcile to the source systems.
- Failures caught before a report goes out.
- Monthly platform cost.
Background
These methods come from And's course plan for ETM 538, Decision Support Systems: Data Warehousing, at Portland State University (Winter 2027).