I start with skepticism. Before digging into the details, I try to understand the whole system: the people using it, the problem it solves and the constraints around it. Strong products require the operational and technical sides to be considered together: architecture and usability, governance and efficiency, cost and security.

I build for resilience. A product serving a narrow need should be designed differently from one serving a broad audience, and neither should depend too heavily on a single model or vendor. I keep knowledge, roles and rules in plain text and treat the model as replaceable, so advances in AI improve the product instead of forcing a rebuild.

I work backwards from the workflow and the decision that needs to improve, not from the model. Before building, I define the user, the available data, the operational constraints and how success will be measured. I use models where judgment is needed, deterministic systems where rules are enough and human review where consequences demand it. Once deployed, evaluations, failure cases and user behaviour guide the next iteration. An AI product is not successful when the prototype works; it is successful when people trust it, use it and the organisation can operate it without its builder.