Writing.

Notes on building AI products, leading product teams, and founder decisions.

  1. What if an app didn't have to fit everyone?

    When customers can change how a product works, a niche request could become something they make and share.

  2. What should an AI tutor make easier?

    Choosing a more familiar reading passage changes the lesson. It also changes what a correct answer can tell us.

  3. How much of an AI product should you script?

    What building a French tutor taught me about deciding which workflow steps an AI can choose and how to check its work.

  4. Every AI product needs a way to change models

    What my first Kimi experiment taught me about model portability, product-specific evaluation, and testing new providers without rebuilding the product.

  5. Where AI should stop in product discovery

    A client investigation about defining the boundary between AI evidence gathering and product judgment.

  6. Why I gave an AI agent a manager

    Why I separated a user-facing AI agent from the operating layer responsible for current state, evaluation, and closure.

  7. An AI agent can be wrong before it answers

    Why evaluating a complex AI system means checking the path through context, tools, state, and actions, not only its final response.

  8. The AI answer that felt right was wrong

    How a plausible, repeatable AI analysis led us toward the wrong model, and what changed after we inspected the query.

  9. When an AI product decides what deserves your attention

    What I learned from defining success for an AI product designed to notice important issues before a product team did.