BuildersLog

Most software relationships end at go-live. We built the company around the opposite bet. Fifteen years in, two platforms deep, we’re still here — which was never the strategy. It was the whole point.

A predefined model ships fast — until the second category of data arrives, and the third. We refused to build Ark that way, and paid for it in the slow early months while the foundation took shape.

The remediation contracts were there for the taking — large, repeatable, predictable. We stopped pitching them. Not because the work was hard, but because winning it meant selling the disease as the cure.

The original pipeline worked. The cost was fine. Nobody asked for a rebuild. But depending on a single provider for a core capability is a decision that compounds in ways you don’t feel until it’s too late.

Everyone debates which AI model to use. Almost nobody talks about the retrieval layer that determines what the model actually sees — and why that engineering matters more than the model itself.

A generous context window lets you be lazy and still get results. An 8K token limit forced a 94% reduction in wasted context — and revealed that better engineering, not a bigger model, was the answer all along.