HERMOD SYSTEMS

Agentic AI consulting and development

Software built by governed AI teams.

The human writes the constitution โ€” and is the one who can be pointed at when it is wrong.

We take on consulting engagements and build projects end to end. Hermod Field, launching October 2026, is one we built for ourselves.

The gap

Everyone is shipping faster.

Almost nobody is shipping further.

Failure 01

Laid by eye.

No square, no plumb line, no drawing. The course looks fine until it does not.

Asked forActually built

Failure 02

Written in dust.

The rules existed. They lived in a chat window, and a footprint went through them.

Not in gitGone by Friday

Failure 03

Three places, disagreeing.

Every scrap is someone's truth. None of them line up, and one is already falling off.

DriftNobody checking

So

The bottleneck is no longer intelligence.

It is that nobody designed the institution the intelligence works inside.

Condition 01

The record has to outlive the session.

What the team knows has to sit where no agent can quietly revise it and any human can read it. We use git. Yours might not. The condition is not negotiable; the store is.

State lives outside the agentsReadable without us

Condition 02

Whoever judges cannot be whoever built.

Independence has to be structural, not asked for politely. We buy it by holding the two seats with models from rival labs. That is one way to get independence, and it is not the only one.

Structural independenceNobody grades their own work

Condition 03

Something has to be trying to break it.

Verification that is nobody's job does not happen. We stand a fresh-context adversary against every build and let it overturn the arbiter above it.

Three conditions, one implementation. Yours will not look like ours, and working that out is most of the job.

Adversarial by designBefore production, not after

How we work

Built by this method ยท launching October 2026

Hermod Field

A campaign field platform โ€” routes, door logs, live results and GOTV, one system the whole field team runs on.

Offline-firstIn the field in October

The Robot Government Diaries

We write down exactly how.

Every failure above is one we hit ourselves. The ones we have written up are in the diaries, dated.

Read the diaries

What we do

Consulting engagements and end-to-end builds.

Governance setup

PLACEHOLDER — one sentence describing what a client buys. Not written yet, and not invented.

Build

PLACEHOLDER — one sentence describing what a client buys. Not written yet, and not invented.

What we built

Four systems, and what can be checked on each.

One line for what each one is, one line for what can be checked from outside. Two of the four can be checked today; the other two say why they cannot. None of it is offered as a client reference.

The human in the loop

Ben Zhou

One person, in Burnaby, British Columbia. I write the rules the AI team works inside, and rewrite them when they crack.

Scratchboard portrait of Ben Zhou, white line on black, signed benz.

About fifteen years in competitive athletics. About a decade writing code. The overlap is smaller than it sounds and more useful than it looks: sport is where you learn that the referee cannot also be the player, and that a rule nobody enforces is not a rule.

So the part I am careful about is not the model. It is the institution around it — who holds which seat, what is written down, and who is allowed to say no. The conditions on the method page are the ones I hold my own projects to.

Day to day that is the job. I set the mandate for each seat and I read what the seats argue about. When the argument shows the rules were wrong, I change the rules. The code is written by the seats.

A note on the word we, since it is all over this site. It means one human and a set of governed AI seats. Nothing here runs unattended, and there is no autopilot claim behind it.

PLACEHOLDER — track record. Employers, titles, dates and any shipped-at-scale numbers are not published here, because none of them has been verified for this page. What can be checked is the work above.

LinkedIn

From the diaries

What we are learning, written down as it happens.

The Robot Government Diaries · Entry 01

The Accidental Government

Everyone keeps asking which AI skills to study. My honest answer is neither a skill nor a model. It's governance โ€” the engineering of alignment.

Read it →

Start a conversation Read all entries