Independent data architect
Models don't invent a reliable world. They succeed running on the data you already have - or they fail on it.
I do the ordinary work. Data warehouses, reports, Fabric bills. I work by the hour. A fixed fee when the job is clear.
The actual work
Finance and sales both call it revenue. The board gets two numbers and a fight.
Open the model both reports claim to use. Not a workshop. The model.
One definition. One path to the report. The argument ends because the number only has one place to come from.
The dashboard is signed off. The number that moves the week still arrives as an attachment.
Ask for last week's decision. Open what they actually used.
The report and the decision read the same thing. That file is the source. The dashboard is the copy.
The Fabric bill went up. The picture of the platform did not.
Last month's capacity. Which workspace, which refresh, which loop.
A short note you can run next month. Keep, stop, or leave it - with a number.
You bought the hands. The eyes are still looking at a mess.
Pick one action. Ask what it would read. If the answer is "we'll clean it later," stop.
Real data in. One allowed action out. A person for the ugly case.
The agent can draft the change. The yes still lives in someone's inbox.
Name the exception. Who can stop it. What happens when it is wrong.
One job. A named stop. A person who can kill it.
Cases
A maritime operator, a capital fund, Danish agencies.
Maritime
Operations, fleet, safety and commercial reporting on Microsoft Fabric. Lakehouse, semantic models, Power BI. A release path so a change could land. Agentic work sat on that same path: models, AI-assisted review, a small internal agent, sources that already worked. The platform is what they run.
Microsoft Fabric · Lakehouse · PySpark · Delta Lake · Power BI · Azure DevOps · Git
Capital fund
Several entities, one portfolio. Enterprise data architecture and the BI on top of it. A Power BI consumption model the domains were supposed to share. Azure SQL, snapshots, bi-temporality, master data. DevOps and tests in the pipeline, so the model was not just a picture in a deck. They still report from that layer.
Azure SQL · Power BI · semantic models · DevOps · CI/CD · automated testing · Master Data Management
Public sector
A shared Azure analytical platform for several Danish agencies. SAP, finance and operations into Synapse, Data Factory, Databricks and Delta Lake, with Power BI on top. ISO 27001. Infrastructure as code, and a DevOps path that had to outlive one team and one agency. They still share that platform.
Azure Synapse · Data Factory · Databricks · Delta Lake · Power BI · ISO 27001
About
I am an independent data architect in Denmark.
Fourteen years across public sector, financial services, media and maritime. MSc in IT Management from CBS.
Agentic development
First the numbers have to add up. Then an agent can act.
Named tables. Named systems. That is the read path. Not the whole warehouse. Not last year's extract. If you would not sign the number, those rows stay closed.
What can be decided without a person, and what still needs one. Write that line down. The next person should be able to find it.
The one thing that may happen - send, book, change, file - and how a person stops the job. One action. A stop someone owns.
Send the break, not a deck. I reply. Then we look.