Data platforms, done right
Governance, permissions, lineage, and the foundation that decides whether everything built on top of it holds up. Notes from working on Databricks, Unity Catalog, and Microsoft Fabric.
There's a lot of noise about AI these days. But if you can't explain how a system works when a real person uses it on a Tuesday morning, you don't really understand it. I write about the practical stuff in the middle: making models, agents, and data platforms actually run, and fixing the unexpected quirks along the way.
I keep the list small on purpose. These are the areas I've actually built production systems in — not dabbled, not read about.
Governance, permissions, lineage, and the foundation that decides whether everything built on top of it holds up. Notes from working on Databricks, Unity Catalog, and Microsoft Fabric.
Generic models give generic answers. Notes on building AI that understands a specific domain's logic, language, and edge cases well enough that the people who own that domain actually trust it.
Agents that do work, not demos. Lessons on guardrails, evals, human-in-the-loop review, and monitoring, the unglamorous parts that decide whether an agent survives contact with production.
Reflections on shipping AI inside a large organisation as an individual contributor and technical builder, and on what it actually takes to grow into leading it.
Why benchmark scores don't make an AI/BI tool trustworthy. Notes on schema design and business-logic standardisation, the unglamorous work that decides whether a Genie-style tool gives a right answer or a confident wrong one.
I started in finance before moving into data science, so I tend to think about AI in terms of what it actually changes for the people using it, not just whether the model works.
This site is where I write that down. Views here are my own, not my employer's.
— Banu Tennakoon
If something here was useful, wrong, or worth arguing about, I'd like to hear it. Always happy to connect with other people building in this space.