My career looks like three different jobs. It is one habit. Find the work people do by hand, understand it well enough to know what is safe to remove, then build the system that does it. The tools changed from Python scripts to report engines to AI agents. The habit did not.
Ran more than 50 process-improvement projects. On my own initiative, I replaced the plant’s paper data-collection system with a mobile app and Python scripts. The plant had real-time process data for the first time. The same period produced a US patent.
The lesson I kept: the hardest part of automation is never the code. It is understanding the process well enough to know what is safe to remove.
Moved from the United States to Canada and studied business administration, finishing with a 4.0. I made the change to software on purpose, and I brought the business half with me.
It is why I ask who uses a system, who pays for it and who carries the risk before I ask how to build it.
A platform used by municipalities and engineering consultancies across Canada, the US and the UK. I learned the domain by building the documents customers actually read: inspections, payment certificates, tenders.
I built the layer around the model: tools, guardrails, memory, evaluations and an autonomous work loop. Then I ran real production work through it. One policy file compiles into the permission gates of two agent runtimes. Everything irreversible stays with me.
Since mid-2025 I have used side projects to learn what my day job does not teach, and I build each one to production standard: ChiOS, the app I run my life on, which agents reach through a 200-tool MCP gateway; chi-harness, a personal agent harness where safety lives in code; a desktop search tool that runs a local model; and a Go authentication library.
Some are deployed and I use them every day. They exist for skill advancement: new languages, edge platforms, security done properly, and agents as users of software. Building for real use is where the learning is.
Harnesses, MCP servers, policy-as-code, evaluations and telemetry. The model is a part that will be replaced. The harness around it is the asset.
Data model, authentication, API, web, mobile and deployment. I pick the language the problem wants: TypeScript, Go, Python, C# or SQL.
Four and a half years of bespoke engines, integrations and a commissioned app for municipalities and engineering firms. Comfortable reading a client’s spreadsheet macros, in English or French.
Chemical engineering and process control taught me to respect the process before changing it. A US patent and a peer-reviewed paper came out of that work.
“Anything you actually require must be enforced somewhere the agent’s reasoning cannot reach.”
“A silent no-op reads exactly like a pass.”
“Every rule sits on a rung. One observed violation promotes it one rung.”
“Evidence before diagnosis.”