Humangentic projects

What I build after the kids are in bed.

By day I write about how organizations run when people and AI agents share the work. At night I find out whether I am right.

These are volunteer projects for my own community: production systems with real users, built by me and a team of agents I direct through ZAMARA, the harness I designed. I stay hands-on because the lessons come from being the person the agents hand their work back to.

Every system here is built for daily use by people who never asked for AI: a parish office on a Monday morning, a parent at bedtime.

Humangentic is my name for the fourth stage of AI maturity, where human judgment and agents share one job and each does the part it is good at. These projects are where I practise it.

I decide

Scope, what ships, and anything that touches money, children or a parishioner's data. No agent pushes to production.

Agents do

Implementation, tests, and review in independent lanes, where one model's work is attacked by another that did not write it.

Gates hold us both

Rules written as code, not as prompts: accessibility checks that fail the build, story language a model cannot talk its way past.

A parish website, and the office behind it

A bilingual site for parishioners from 13 to over 65, and a staff console for the people who run the office.

The brief

St. Faustina is a Roman Catholic parish in Cumberland, Ontario. Its previous site ran on a church website builder and could not do what the office needed. The people it serves sit at two ends, parishioners over 65 and young people from 13 to 30, in English and in French.

What shipped

A public site covering Mass times and the weekly bulletin, the livestream, sacraments, parish groups, hall booking, giving and pledges, the building campaign, events, and a youth hub with its own look. Behind it sit the parish workflows: the office's manual activities, upgraded to digital workflows. And print views, because a parish still runs on paper.

How the agent team ran

The team ran inside ZAMARA, and its first rule is that no agent marks its own work. Code written in one lane is attacked in others: three Claude lanes and Codex as an outside peer, each working independently. Thirteen four-lane reviews and five three-lane reviews are on file, and one of them ran 34 agents. Before a job, each agent reads the lessons file for its role. After a fix, it adds to it. Live behaviour is checked in a real browser from a written runbook.

The catch worth telling

More than three weeks before go-live, an independent review lane found a critical flaw in code that was already written and in the repository. It was fixed and pinned with a regression test. The agent that wrote the code was never going to find it, because it was the one that believed the code was right. That is the case for adversarial lanes.

“Never delete a contradicted lesson.”
Standing rule in every agent's lessons file

With Sammy: stories that meet a child before the hard moment does

Free, open-source social stories for autistic children. Named for my son.

The brief

A social story walks a child through a situation before they meet it: visiting the dentist, the first day of school, a busy brain at bedtime. We went looking for help and found therapy waitlists that stretched for months and apps that felt clinical and cold. With Sammy is free for every child, in the words of its own mission, “regardless of their family’s income.”

What is built

Sixty-two stories, each in three forms (full text, simplified, and symbols only) so a child who does not read or speak still gets the story. Three languages. A story builder for parents that starts from the reason behind a behaviour (escape, attention, access or sensory) and not from the behaviour alone. Communication symbols from ARASAAC and Global Symbols, read-aloud, and offline use.

Where a model gets no vote

Social stories have a standard, Carol Gray’s Social Stories 10.4, and it is strict about tone: describe, never command. Here that standard is code. A story containing “should”, “must”, “ought to”, a “you”, a threat, a bribe or shame fails the gate, whatever the model thinks of it. Every generated illustration passes a child-safety check before a child can see it. The site collects no personal data from children.

How the agent team runs

Inside ZAMARA, work moves through a fixed loop: one agent implements, a second verifies, a different model reviews, and fixes go round again. Audits use six independent reviewers who must cite file and line for every finding. I merge every change myself.

“Never weaken a clinical rule or accessibility guarantee to make a test pass.”
Standing rule in the hardening plan, July 2026

Still to do before launch

Clinical reviews and licence approvals
Age-graded audio narration for children with visual impairments
Illustrations tailored to each age band across the story library
Optimize accessibility tests to meet the standards

Running something for your community that needs building? Tell me about it.