I make complicated systems tell the truth about how they work.

The working method

Four moves I keep finding in the work.

1. 01

   Walk the land

   Read the system, talk to the people inside it, and observe what it does before deciding what it needs.

2. 02

   Map the real bounds

   Observe the system directly, change one condition at a time, and map where reality stops matching the dashboard.

3. 03

   Complete the circuit

   Connect the useful prototype to traffic, operations, recovery, and the people who must decide what happens next.

4. 04

   Leave a trail

   Write down the evidence, the wrong turns, and the operating decisions so the next person does not start from folklore.

Much of what I learn comes from ordinary work: an unexpected result, a small adjustment, or a question from someone else. I follow those threads and keep notes I can return to.

2005–2026

The tools changed. The questions did not.

1. 2005

   The hidden field

   I was learning to pull component descriptions out of CAD assemblies through the SolidWorks API. In a public debugging thread, the useful correction came down to the case of a letter. Make the failure visible, accept the correction, keep going.

2. 2012

   A world you could walk through

   I built an interactive simulation by hand. It reinforced something that still matters in production software: a model becomes useful when people can move through it, test it, and see where their assumptions stop matching the world.

   That year, I also wrote a custom HTML5 canvas particle animation for a nationwide mobile carrier rollout. I remember the turnaround as 72 hours. The retained source shows a shared timer driving motion through tunable multipliers—a design I approached through the base-clock-and-multiplier model of PC overclocking. The work was about fitting the animation to the mobile hardware and making its behavior visible enough to tune.

3. 2018

   The factory floor

   I delivered a greenfield order-entry platform for a supply chain spanning more than 100 factories and $300M in yearly order volume. The less visible work was just as important: telemetry, operating help, deployment, and a team able to own the system after handoff.

4. 2019

   Scarcity gets a model

   I wrote about job-shop scheduling from inside manufacturing. Seven years later, the same constraints showed up while scheduling AI agents and GPU model loads. Different floor, same discipline: the solver is only as trustworthy as the observations underneath it.

   My 2019 archive also includes Clean Architecture and CQRS reference material exploring separate write paths, read models, and append-only event histories. Those questions—where truth lives, how a view is derived, and what history is retained—continue through my work on agent systems and replayable state.

5. 2020–21

   Complete operational circuits

   Led customer-identity consolidation (CIAM) and Auth0 externalization across multiple engineering teams for a consumer-credit platform handling 100M+ daily API requests and eventing for billions of monthly identity events. Identity, APIs, security, telemetry, testing, documentation, and cross-training were not separate deliverables. Together they made systems that could be operated, changed, and defended by the people responsible for them.

6. 2026

   The instrumented workbench

   Local AI infrastructure, agent control systems, game networking, and causal hardware investigations became a connected body of work. The hardware campaign separated three distinct bottlenecks; the Vulkan investigation moved the actual software wall experimentally.

   Farmer’s architecture makes the filesystem the source of truth: each run keeps its request, state, events, and results on disk. The record remains available for inspection after the process stops. It is the same habit of making behavior observable, applied to an agent system.

   Lumberjacks started with an AM4 first pass. The work took weeks of difficult integration and multiplayer testing. I completed the final cutover fully remotely on GCP, driving OMEN and i5 as players. After validation, I rolled AM4 back for standalone mods that lowered adoption requirements and widened tester participation, and rolled back or shut down GCP to control costs while building the surrounding tooling ecosystem. Read the cutover and wind-down history →

   Those integration-tested building blocks made later ideas quicker to try. DMos reached a friend-tested PoC in under 24 hours by connecting existing spatial, game, and local AI work through a new interface.