Infrastructure, fabrication, systems

Built.
Earned.
Understood.

Hi! I'm Louis!
This is my page! I'm a husband, a father, a savant of fabrication and welding, and I think I've done some cool stuff. Ironman 140.6 Finisher. Probably quite Autistic.

See the work →
Louis Poulin at Niagara Falls

Previous Employers

These are the two most respected companies I have worked for, and I thank them explicitly for the opportunity to serve.

Education

I graduated high school. A few years later, I earned a 4.0 in Welding 1 at Lansing Community College. Fourteen years after high school, I earned a 4.0 in Precalculus at Lansing Community College.

Michigan Stadium structural steel installation

Michigan Stadium Project

Michigan 2023 National Champions badge

Construction of the expansion began on November 19, 2007. I went on to work with both hired fabricators to cement my legacy to that job. Jim Harbaugh is a personal hero of inestimable worth. Nothing I've worked on before or since has scratched that itch, and I've been looking for it my entire career.

Michigan 2023 National Champions badge
Flame hardened 4140 blade result

4140 Flame Hardening Recovery

An ad-hoc flame-hardening recovery on a 4140 blade, built from a broken procedure I was trying to reinvigorate. The final product exceeded standards and hit Rockwell 59 consistently.

Work roll rail straightening setup

Work Roll Rail Straightening

The welders didn't account for warping, and I had to straighten a 2 inch curve and 1 inch camber out of a work roll rail. It took multiple stress-relief passes, plus pre-machining weld build-up. Nightmare fuel.

Top wrapper new build

Top Wrapper Build

A new top wrapper build, the first one made there in 40 years. It was machinist verified to no more than .006 global off tolerance. Paul was the engineer on it, and he told me I was the most talented man he'd met in the last 20 years.

Weld build-up for machine finishing

Lathe Bearing Diagnosis

Weld build-up for machine finishing, with distortions visible in the weld pattern. The bearings on the lathe were bad, and they blamed the rookie foreman for not knowing how to weld. Guess who was laughing when we had to emergency pull and replace the bearings for a 1955 German lathe.

Gallery
Selected

Front-page images are controlled by photos/front-page.json.

Louis Poulin

Louis
Poulin

Infrastructure Engineer · Builder · Athlete

I build and operate infrastructure for a living and for sport. The lab runs a fleet of Proxmox hypervisors, TrueNAS storage, a multi-vendor GPU stack, and more ZFS pools than any reasonable person would want to maintain.

When I'm not in front of a terminal, I'm usually in the workshop or moving under my own power somewhere. Everything here is documented the way it deserves to be — not the way that's easiest to post.

Proxmox ZFS Self-Hosted Endurance Fabrication HV-Katra Michigan

The Lab

A machine shop for ideas that haven't been proven yet.

I built it to answer one question: how useful can modern AI become when it is connected to evidence, real systems, and disciplined engineering practice?

My background isn't software. It's manufacturing. Decades of building, fitting, welding, troubleshooting, and improving physical systems. Steel, machinery, production processes. That work taught me three things that don't change regardless of the material:

Reality

Evidence matters more than opinion. The material tells you what it is, not what you hoped it would be.

Reproducibility

If you can't reproduce a result, you don't have a result. You have a story.

Failure

Every failure teaches something. The ones that don't teach anything were poorly documented.

Procedure

Good systems are built on tested procedures, not on assumptions about how things ought to work.

When the AI era arrived, I didn't treat it as magic. I approached it the same way I approached manufacturing: build a shop, develop procedures, gather evidence, test assumptions, and learn how the machines actually work.

That shop is what you're looking at.

// current research

Active Projects

Some succeed. Some fail. All of them leave evidence behind.

Operational Tooling

Agentctl

An experiment in AI-assisted operational tooling. The question isn't whether AI can write code - it's whether AI can operate a system responsibly, on evidence, over time.

Active research
Verification

Evidencectl

An evidence collection and validation framework. A system that cannot prove what it did is a system that cannot be trusted. This is the foundation everything else runs on.

Active research
AI & Reality

Reality Contact

Research into maintaining trustworthy relationships between AI systems and real-world observations. The goal is grounding, not performance.

Active research
Media & Compute

Arc Pro B70 Research

Ongoing work investigating AV1 encoding, media pipelines, and reproducible performance testing on Linux. Xe2 is interesting hardware. The driver story is still being written.

Ongoing
// why publish it
"I believe individuals can do meaningful technical research. You don't need a corporation. You don't need a university. You need curiosity, discipline, and a willingness to publish what you learn."

Most technical websites about AI were written by people who have never debugged a failing weld or figured out why a production line stopped at 2am. This one wasn't.

The rarest thing in any field isn't intelligence. It's someone willing to show their work before they know how it ends.