Principles.

Before looking at existing solutions, we write down what we would build if the field did not exist. This takes longer upfront. It pays back every time the field changes.

We do not build foundation models. We do build every layer between a foundation model and a production decision. Those layers are where our judgment lives.

A simple system run with discipline compounds. A clever system that requires constant attention does not. We pick the simpler design when in doubt.

An idea that only works in a notebook is not a result. We do not publish anything until we have operated it in production — however small that production is.

We spend more energy on retrospectives than on roadmaps. The roadmap is a guess. The retrospective is data.

If a decision is not written down, it will be relitigated. Writing is not overhead. Writing is memory that compounds.

We optimize for the shape of the system in five years. Not five weeks. Some decisions that look bad in five weeks are correct at the five-year horizon.