The idea
Plants don't have a data problem. They have a capture problem.
The richest source of operational truth — what operators see, hear, and do on shift — is rarely captured in a form any system can use. It's spoken over radios, held in memory, scattered across spreadsheets. And lost.
I call it the capture layer — turning what the floor already knows into structured data, at the moment it happens. Not bigger models. Better capture.
Capture at the moment
The watch-item, the half-finished job, the warning at handover — logged when it happens, not reconstructed after.The model never decides
Language captures and classifies; a deterministic engine stays the system of record. A design an operating plant can trust.Built on the floor
Designed from the shift, not from a spec sheet. It works because the person who built it has stood where you stand.The full argument, in one piece — The Capture Layer →
The shift log — writing
The Capture Layer
The whole thesis in one piece: what capture actually is, the three ways plants leak it, and why "we'll do AI later" quietly fails.
Read →High Vibration Is Not a Severity
A model kept reading "high" as urgent. On the floor it's part of the fault's name — and that gap is why floor knowledge beats model scale.
Read →When the Open Item Carries Itself Forward
An open item survives a shift gap one way — someone re-writes it by hand. Re-writing is the failure mode.
Read →Close the Loop, Not the Log
Mineral processing was never a slow industry. So why did logs, tickets and handovers stay frozen?
Read →Noticing Was Never the Problem
The warning gets picked up. What a shift can't do is hold it all in its head until morning.
Read →The Floor Has the Edge
The people closest to the problem hold the data nobody ever wrote down.
Read →What a Handover Has to Carry
The four things every shift change has to move — and what breaks when one slips.
Read →Piper Alpha, 38 Years On
What the gap between two crews can cost, and the handover discipline that closes it.
Read →The latest entries are published here in full. Pieces first appear on LinkedIn — full versions land on this page.
Learning in public
Alongside the writing, I'm teaching myself to build automations that take the manual busywork out of operations — reports, summaries, the paperwork around a shift — and sharing what I learn, in plain terms, for other operators. A few of the things I'm building:
Refinery OS
A shift-intelligence platform for plant floors: capture every equipment defect and shift handover the moment it happens, so the history is always there. I designed it from thirteen years on shift, and run it myself.
Golem's Docs GPT
Grounded Q&A over your own documents — every answer traced back to the exact source line, never invented. Built on Claude; bring your own key.
View on GitHub → Open sourceGolem's Brief Deck
Turn any long video into a clean research report or a navy-and-gold briefing deck — the reading I never have time for, done in one pass. Open source, bring your own key.
View on GitHub →One spine, many floors
The same capture spine has been cloned for marine operations and training — days, not months. The idea isn't refinery-specific. It's shift-specific.
Buildings run shifts too
A tenant's 11 p.m. message is the same uncaptured signal as an operator's radio call — after hours, unstructured, gone by morning. The same capture layer, applied to property.
Read →One thread
Every lesson here is the same idea in a different coat: catch the signal the moment it happens — before it's gone.
And the part I care about most — what building these actually taught me. Every one started as a mistake or a surprise on my own screen, and the floor taught me to write those down. Full versions land on LinkedIn — the short of it:
Test the road, not the map
This site looked broken to every real visitor for an afternoon — while every check I ran said it was fine. My tools kept fetching fresh files; a returning browser was holding four-hour-old ones. Now I verify the exact path a reader walks, not the one my tooling takes.
The error that never fires is the expensive one
An app I was building was quietly dropping data fields — no crash, no warning, just gaps you'd find months later. I built a guard that catches it before every release. It's the whole reason I build capture tools: the miss you can't see is the one that hurts.
Make it deterministic before you make it smart
I put the AI step last on purpose. Resolve the equipment by plain rules first; let the model help only once the data underneath is clean. Point an AI at a messy foundation and it doesn't fix the mess — it multiplies it.
A load test is insurance, not proof
I ran 457 writes to be sure the app wouldn't fall over. It didn't — and that proved nothing about whether the thing actually solves the problem. Real proof is one person doing the real job with it, once.
The skim is the first read
I wrote a piece I was proud of. The verdict came back: "a boring page." On a screen, people scroll the whole thing in ten seconds before reading a word — so the headers and the diagrams have to carry the argument alone. Rebuilt for the skim.
Four AI sessions, one permit board
I run several AI assistants at once, and they kept colliding on the same files. So I gave them a permit-to-work: claim it before you touch it, and the others see the claim. The floor discipline that keeps two crews off one pump, applied to code.
About
I'm not a software vendor who read about plants — I've run them. Thirteen years on the floor: gold processing in the Philippines, then alumina refining — the Bayer process — at a Gulf refinery, as a licensed metallurgical engineer and shift supervisor who issues the permits and leads the crew.
I build the tools I wished I had at the 6 a.m. handover. Whatever I make works because the person who made it has stood where you stand.
Stay in touch
Get the next piece.
Get the free field guide — The Handover Gap: three disasters, one lesson, and a checklist for your next shift. Plus the occasional piece worth your shift. No noise, no pitch.
Run a shift somewhere? I compare notes with operators and supervisors on how real floors handle handover and capture — what works, what quietly doesn't. No pitch; I'm not selling anything. Tell me how yours does it: [email protected]