Notes from building and running production AI for companies. Many of them start with something of ours that failed. Each one carries a name. We promise no schedule and send no newsletter: a note goes up when there is something worth writing down.
Every note, by month.
October 2026
- Approvals that continue the workSaying yes used to leave the work stalled, or bring the same question back. What broke after the click, what we changed, and what is still not right.712proposals that expired unanswered across our workspaces in 40 days, against 97 that ran
- One ask, one runSomeone asked our product for a deck and got four runs, three approval cards and no file. What was wrong with each piece, and the run that delivered.11minutes from one ask to a finished deck, with zero approval requests
September 2026
- A skip on a critical alarm is a silent failureTwo critical health checks answered skip for days while questions to our brain failed. The board showed nothing wrong. What we changed.1,791skips in a row from one critical check, over 12 days
- Every sentence the model wrote was trueOur assistant told a user it had taken them to a page and moved nobody. The model was not making things up. The system around it was wrong.38tokens the navigation tool missed the cut by
- A run waiting on you is not a hung runOur plan card asked "stalled?" while a teammate was working. Then our own production test called a correctly waiting run a hang. What we changed.7states a quiet run can now be in, and only one of them is stalled
- Broken and empty are the same pixelsA failed read that renders as an empty state tells a client something false. Where we found it in our own platform, and the checks that now stop it.5client screens that mishandled a failed read, fixed on 6 August
- We drew our icons instead of picking a libraryOur product rail used eight icons for more than one thing. We drew 31 of our own. The first round looked right as numbers and failed at 16 pixels.31icons drawn for the client rail, each with an outline and a solid state
June 2026
- RAG is not training: a plain-English guide for ownersRetrieval and fine-tuning solve different problems. A buyer's guide to RAG, hybrid search and rerankers, and the vendor phrase that should end the meeting.
- The enrichment layer is where outbound systems dieAnatomy of a real reply-to-deck pipeline: reply detection, research, three-stage strategy, decks and a Slack approval gate, on one reconciled record.
- What an honest AI security page looks likeWhy we publish what we are not certified in, the claims an AI firm can make honestly, and how a buyer can verify every line without trusting us.10risk categories in the OWASP Top 10 for LLM applications (OWASP, 2025)
- Why we push complexity into deterministic codeFive 90% steps compound to 59%. How the DOE architecture keeps judgment in the model, control in deterministic code, and a person at the gates that matter.59%success rate of five chained 90% steps (0.9 to the fifth power)
Run your whole company on AI.
Tell us how your business runs. If we're not the right fit, we'll say so on the call.
Your custom strategy and systems, delivered in under a month. Your named people using them every week by day 30. Or we keep working for free until both are true.
The AI operating layer for enterprises.