Presentations ============= **Source Control in Power BI: a BI developer's journey to CI/CD**, Athens Fabric & Power BI Meetup, March 2025. A talk arguing that Power BI development deserves to be treated as software engineering, and that Microsoft's new project format finally makes it possible, opening a traditionally closed, proprietary Fabric environment up to version control, review, security, and even AI agents. I built and delivered it the way I build documentation: a structured, sourced case aimed at moving a community of practitioners toward a more open and accountable way of working. What it covered: - **The problem.** BI developers are "software developers supercharged with business knowledge", yet were stuck editing binary ``.pbix`` files, versioning by saving copies, and exporting to Excel, with no diffing, no review, no safe collaboration. - **The grounding.** I set the argument on the business-intelligence-systems literature (Davenport and Arnott; the finding that roughly 70% of BI projects fail to meet expectations without proper foundations) and closed by mapping the approach to the DeLone and McLean information-systems success model. - **Microsoft's "revolutionary trinity".** The PBIP structure with the **TMDL** text format (turning binary models into transparent, version-controllable, human- and agent-readable folders), Git integration with Power BI Service workspaces, and Copilot integration. - **A metadata-only security model.** Models reference data sources by name only, so no production data ever travels inside a file, which satisfies strict compliance requirements. - **The real implementation** at Maya: a shared "universal" core model, automated workspace and customer-folder creation via the Power BI API and GitHub Actions, a branch โ†’ test on a dummy workspace โ†’ pull request โ†’ merge loop, a standardized ``.gitignore``, and Python automations (for example, rebinding report visuals straight from ``tmdl`` metadata). - **An agent-experience outlook.** Because source-controlled Power BI artifacts are granular, structured text (JSON plus ``tmdl``), agents can parse and even generate reports without proprietary connectors. `Slides `_ ยท `meetup page `_