I run operations. I build the systems that run them.
With AI coding agents, on a locked‑down desktop. The order platform I directed went into daily production in September 2026.
From the warehouse floor to software.
I trained in physiotherapy and moved into operations in Canada: office operations first, then a warehouse team I led on SAP, and since 2024 logistics, warehousing, daily operations and customer service at one branch of a publicly listed commercial-equipment company in international trade.
I started with AI in 2025 and built my first tools inside spreadsheets and the browser. In August 2026 I put an AI coding agent on the work machine. I write the specification, decide the rules and require measured proof. The agents write the code.
An order desk that keeps the ERP in charge.
A dealer's purchase order is read twice, checked line by line by a person, and sent to the ERP as a draft. The ERP's own approval rules decide what becomes an order. Pick lists, shipping documents and one archived PDF per invoice follow.
What changed for the branch.
My own estimates from daily operation, not instrumented measurements.
Problems I solved, and how I proved it.
Each one changed how the system is built, or how I run the agents.
Measure the theory before building the fix
Everyone suspected the file server. A canary and a few counters pointed at our own service, and the fix worked on the first try.
Read Story · October 2026 18% → 64% whole orders read exactly rightThe ERP as the answer key
Nobody had time to label data. The filed paperwork already paired every purchase order with the order the ERP booked. Whole-order accuracy went from 18% to 64% in nine days.
Read Story · October 2026 13 / 25 → 39 / 0 door tests passing / failingOne action, one control, one logic
A fix added a second way to put a file into an order. I made a rule that a second way is a defect, and the agents proved it door by door.
Read Story · October 2026 44 / 95 min time left against the full test run“Not released” can be the gate working
A small, well-proven change, a forced logoff, and 44 minutes against a 95-minute test run. The agent reported not released. That was the right answer.
Read Story · October 2026 269 / 269 doors and menus passing, 7 role shapesProve it as every role
A supervisor's order board hung on “Loading”. Every proof had run as an administrator.
Read Story · November 2026 167 facts measured in 47 minutesMeasure the new machine before touching it
The workstation I built on was shut down at a few hours' notice. Two of my own claims about its replacement were wrong within a day, so a probe measured it instead.
ReadHow I work with AI agents.
The method grew one work packet at a time. Each rule exists because something went wrong without it.
Three roles, fixed
An architect model writes each work packet. A coding agent builds, tests and releases. I decide anything that touches money, permissions or production.
Outcomes, not steps
A packet states the goal, why, what done means and every red line with its reason. The agent owns the how.
The cost of a mistake picks the model
Four run classes, from rule-settled work on a mid-tier model to judgment work on the top tier. A class can go up, never down.
A different model reviews
The reviewer reads, writes the fix, and hands it to a cheaper model to carry out. Then it checks once more.
Proof that can fail
Red before green. A positive control behind every “nothing found”. Every role, on the installed build.
One release per packet
Every suite green, a signed bundle, every desk updated within the hour. “Not released” is an allowed answer.
Know the bill
Cost is context × turns × agents. 89–96% of tokens are cached context, so context is cleared and long output goes to files.
One home for every rule
Each rule lives in one place with its source. A rewrite moved 171 rules with a script proving each one arrived.
Seven agent skills I wrote.
A skill is a pack of instructions an agent loads when a task needs it. Mine hold 72 anti-patterns and 40 written lessons, so every run starts from what the project already learned.
Platform change protocol
The work-packet contract, definition of done, release and hand-brake rules.
ERP and warehouse data access
Four answer states, probe columns first, a missing price is never zero.
Locked-desktop deployment
Per-person services, watchdogs and signed self-updating bundles without admin rights.
Tool-building standard
A pre-build checklist, the acceptance triangle and 72 anti-patterns.
When to build an agent
Four tests before deploying one, and capturing expertise where work happens.
Rulebook update
A weekly update of every rule, each with its source, installed as one package.
Work-board integration
Atomic writes, validation first, and detection for anything dropped in silence.
How I built them
Where they came from, how the rulebook stays true, and how an update ships.
Let's talk.
Open to roles in operations automation, internal tools and AI-enabled operations. Based in Metro Vancouver, British Columbia.
Outside work: audiophile headphones; indie rock, post-punk, experimental and metal; cult and art-house film; game modding and reverse engineering.