It started with a booking problem.
I run a small rental business, koduaed.ee. We needed a calendar and a booking system. The options we looked at, IzyRent, BookThatApp and a few others, had gaps: either missing functionality or no proper multilanguage support. As an IT person I built one myself. That was the plan.
I couldn't stop there. The booking system turned into a Shopify app called HyreKit. I pushed it to the App Store and moved on to the next problem.
That next problem was AI. I had been using AI help during development, but the early versions needed constant correction: wrong context, wrong assumptions, outputs that looked right but weren't. At some point I stopped treating it as a tool and started treating it as an engineering problem. A useful AI needs a canon, a defined domain, and something to be responsible for. So I built that structure, then gave it teammates.
One AI became a board. I do not run the day-to-day work myself. I configure the engine and the AI board, and retain final approval control over money, contracts and releases. Each member owns a domain: strategy, finance, commerce, product, integrity and marketing. The board runs on an engine that manages memory, tasks and a shared knowledge base, built on public course materials and research papers linked on the methodology page; no affiliation or endorsement. The first usable engine emerged in about three weeks over a summer holiday. What runs today is the result of the releases, failures and controls added since.
The changelog on the website has the steps with references to the source materials and what actually got built from each one.
I don't play computer games anymore, but this whole thing feels like the ones I played back in the day, SimCity and Civilization. Running the AI board is something like raising Tamagotchis, except they cite papers and sometimes disagree with my assumptions. I let them.
What started as a small experiment is now a production company: a live Shopify product, a store running on it, and client work under contract. Separately, WatchGuard selected an agentic AI security automation idea I developed for development in its AI Innovation Challenge. That recognition applies to the security automation idea, not to Offsite's engine, board or governance model. Offsite stays self-funded. The open question is no longer whether the thing works, but whether this architecture compounds into a durable company.