Fifty-five hours: a public registry built specification first
I took a public-facing registry from scratch to production in about 55 hours of development time. The specification carried most of the load, and a person approved every handoff.
The CareFirst Practitioner Registry is a public service on alberta.ca. Members of the public use it to find a practitioner, and practitioners use it to enrol. I built it alone, from an empty repository to production, in two weeks of development time: about 55 hours.
It was the main pilot for AI-assisted delivery in our public services, and the pilot succeeded. This post describes the method, because another team can reuse it.
The specification came first
The work started with documents. AI agents drafted the business and architecture documents first: the product requirements, the functional and non-functional requirements, the user journeys, and the architecture. Each requirement cited the government standard behind it, and a deterministic check confirmed that every citation resolved to a real rule.
I reviewed and corrected those documents before any code existed. A requirements document is cheap to change, so this is where the expensive mistakes were caught.
The code came from the documents
Once I approved the documents, the application was built from them under an AI delivery playbook. The agents wrote code against the approved specification, and I reviewed what they produced. The documents stayed the record of what was decided, which is what a reviewer asks for later.
The application runs on the Alberta Digital Service Platform. The platform’s existing services for hosting and identity meant those decisions were already made and already reviewed.
What made 55 hours possible
| Factor | Effect on the hours |
|---|---|
| Specification first | Mistakes surfaced in documents, where they are cheap to fix |
| Standards cited in place | Each requirement arrived with its standard attached |
| Platform services | Hosting and identity came from patterns that had already passed review |
| One accountable approver | No handoff waited for a meeting |
The model wrote most of the text and most of the code. The hours went into reading, correcting, and approving.
What the pilot does not prove
It is one system, built by an experienced architect who already knew the platform and the standards. A team new to both should expect more hours. The pilot shows the method works on a real public service. It does not yet show how the method scales across teams, and that is the next thing to measure.
The registry went through user acceptance testing and a demonstration to the business area before production. AI changed how fast the documents and code were produced, and a named person approved the release.