BLOG

Engineering notes

Field notes on evals, routing and cutting AI spend.

The ledger: splitting "savings" into three verifiable numbers

Potential, realized and verified, and why we only bill on the third.

Most “the model can't do it” problems are really unclear task definitions.

From narrow slices to skills to owned weights to heavier techniques, one rung at a time.

A simple framework for deciding along two axes: cost and risk.

How you score the work is a statement about what matters.

What we learned the hard way about reliability, cost and observability.

Cache misses, model thrash, context bloat and duplicated work.