Resources

Everything I have worked out, written down.

The part of the job that does not fit in a case study: why company AI stalls, what adoption actually costs, and the small disciplines that keep a system trustworthy once it is running. Written from four production builds rather than from a framework, and held to the same rule as the rest of the site. No number without a source, no story I cannot back.

Start here

A reading order, not a feed.

The newest piece is rarely the one to read first. These four are ordered the way I would explain the work to someone who had never seen it: what goes wrong, what it costs to fix it properly, then how the machinery works and why it refuses things.

7 pieces

Adoption Why most company AI rollouts fail The short versionCompany AI does not fail on the model. It fails on two things nobody buys a model to fix: the system never learned the company’s actual facts, and nobody trusted the output enough to send it. A better model fixes neither, and the second rollout dies the way the first one did. 2026-08-11 · 7 min · Essay

Adoption Adoption is the unglamorous half The short versionThe half of the work with no architecture diagram is the half that decides whether anyone opens the thing in month three. 2026-08-12 · 6 min · Essay

Method The intake form is the product The short versionThe first artifact of an engagement is a question set, not a model. Sixty fields, each mapped to the file it produces, and every blank becomes a tracked question instead of a guess. A guess that gets into the facts file comes back a week later as something the company said. 2026-08-12 · 6 min · Essay

Guardrails Two gates over the same content beat one careful pass The short versionTwo independent checks over the same content beat one careful pass, because one pass shares a head with the work and inherits every assumption that produced it. Skip the second instrument and the file that ships wrong is the one somebody had already cleared. 2026-08-12 · 4 min · Field Note

Guardrails A control nobody has watched fail is a belief The short versionThe most dangerous state for a safeguard is not absent. It is present, broken, and reporting clean. Nobody rechecks a green build, which is how the thing the control existed to stop ships anyway. 2026-08-11 · 2 min · Field Note

Method Never ask a client to organize their knowledge. Give them a shape. The short versionNever ask a busy owner to organize their own knowledge. Build every folder in advance and let the structure do the sorting. Asking for organized files gets you an empty shared drive and a build waiting on it. 2026-08-11 · 2 min · Field Note

Talks The room I almost did not get to stand in The short versionThe 2023 Hide In Plain Sight keynote, and why being on both ends of the same story decided the kind of thing I build. A tool somebody can run keeps working after I leave. Work done for them stops when I do. 2026-08-12 · 5 min · Talk

The talk

The room I almost did not get to stand in.

The 2023 Hide In Plain Sight keynote, and why being on both ends of the same story decided the kind of thing I build. A tool somebody can run keeps working after I leave. Work done for them stops when I do.

Nothing loads from YouTube until you press play.

Read the written version