Thoughts on frontend architecture, AI engineering, and building products that scale.
BAML gave Finnbud type-safe AI outputs fast. But as our AI layer grew, prompt management, versioning, observability, and evaluation became more important than the parser. Here is where we kept BAML and where we moved on.
Open Knowledge Format and vector databases both help AI agents access knowledge, but they operate at different layers. One is a format. The other is infrastructure. Here is how to choose.
I was never a fan of Tailwind until I understood why it was created. The debate is not about which is better. It is about which problem you are solving.
After building across Next.js, React Router, and Keycloakify projects, these are the techniques that keep my React code fast, readable, and maintainable.
Most upload components try to own too much. Uplofile exists because upload UI should be composable while your app stays in control of the upload pipeline.
Zustand is tiny and fast until you force Redux habits into it. Use less global state, tighter selectors, and better store boundaries.
One of the most misunderstood fields in package.json is also one of the most dangerous. Misuse it and your library may work locally but fail in production.
Why backend-owned HTML emails did not scale for us, and how moving to React Email with generated types fixed it.
Next.js is not just React with routing. When used as intended, it changes how you think about rendering, state, and UX.
We casually use library and framework interchangeably. They describe different patterns of control. Understanding that difference makes their behavior clearer.