Notes.
Engineering notes on applied AI — honest, practical, written from production. What worked, what didn't, and how to tell the difference before you spend.
How to measure an AI project: the numbers to capture before and after
Five numbers written down before the build are the difference between an AI project you can defend at month three and one you can only argue about.
"Agentic" is a workflow, not magic: AI agents for business, explained
What an AI agent actually is — a workflow with judgment calls and checkpoints — and the three questions that puncture any agent pitch.
An honest AI-readiness assessment for a small business
The five checks we run in a first conversation — data, process, the plain-software test, ownership, risk and money — including the answers that should stop you from buying.
When plain software beats AI (and how to tell before you buy)
Cases where a cron job, a form, or a database index beats a model — and the two questions that tell you which one you're looking at.
The real cost of a custom AI assistant (the demo is 20% of it)
A demo proves the model can answer; the other 80% of the budget pays for knowing it keeps answering correctly — and catching it when it doesn't.
RAG vs fine-tuning for company knowledge: a framework, not a religion
Most teams ask whether to train a model on their documents — three questions show why the answer is usually retrieval, and when it isn't.
Why automations break — and the four properties of ones that don't
Most automations fail the same way — silently, on a Tuesday — and the fix is four unglamorous properties, not a new tool.
What to automate first: scoring a business workflow before we touch it
Start with the boring workflow that costs the most in hours and mistakes — usually intake — and let it pay for the impressive one.