Notes

a correction is flagged, never erased.

Every note is dated. When a measurement turns out to be wrong, the correction appears at the top, and the note that corrects it says what was wrong.

The running series

• 2 notes — latest 10 September 2026 Data

  What has to happen to a corpus before you hand it to anything — and why doing it afterwards is documentation, not control.

• 6 notes — latest 8 September 2026 Self-hosted inference

  What a model costs, what limits it, and what it reveals when you run it yourself on hardware you bought.

• 1 note — latest 4 September 2026 Managing agents

  I manage thirty engineers. On a personal project, several positions are held by agents. Two management experiences, not one plus tooling.

All notes

2026

• Data — September 10 The failure that raises no error

  A scanned document produces text, so a pipeline that expects text works. It also works when it has read only part of the document, and it says nothing.

• Self-hosted inference — September 8 It is almost never the model

  One hundred and nine missed statements, five mechanisms, and the model is behind none of them. The vocabulary gap everyone assumes is dominant turns out to be the smallest category.

• Managing agents — September 4 An agent never asks what I mean

  I lead engineers by day and an organization of agents by night. The difference is not speed. It is that the cost of framing no longer amortizes.

• Self-hosted inference — September 3 A third of my failures were not failures

  I published six values of one metric and wrote that the measurement was wrong. Looking for why, I found a third of the failures came from the instrument — and two unrelated measurements landing on the same number.

• Self-hosted inference — September 2 Why these notes start midway

  The work began in February. Publishing begins today. What this site holds, what it does not, and why the early notes carry two dates.

• Self-hosted inference — September 1 Six values of the same metric

  Recall moved through six values in a week without the model changing. Three causes. This reading has since been corrected.

• Data — August 28 Anonymize before the model, not after

  Stripping direct identifiers is necessary and not sufficient. Whatever has entered the model cannot be recalled.

• Self-hosted inference — August 27 A model that answers badly does not crash

  Errors at zero, latency at 98 ms, no alerts. And the answers are wrong. Conventional monitoring says everything is fine.

• Self-hosted inference — August 25 Six months running a model on my own server

  Three lessons nobody sold me: the hidden cost is not the GPU, your quality metric lies until proven otherwise, and trust is earned through what the model is forbidden to do.

RSS feed