notes.

short essays on what i keep seeing inside companies — workflows that break, projects that stall, and the occasional thing that actually works.

one idea each. published when the thinking is finished — not when the calendar says so.

the problem is almost never the model. it is what sits underneath it.

half of the ai projects i have audited would have shipped more value as a clean script.

the technology is rarely the bottleneck. the calendar, the trust, and the quick win are.

three patterns repeat across every post-mortem i have read.

a one-page decision rule before adding a model to anything.

pick a process. fix it. only then ask what part of it should learn.

measurement is the loop that lets an ai investment survive past q2.

the cost of a new tool is almost never the license. it is the decision overhead.