What we’ve learned shipping AI into real organisations: method, governance, and the distance between a demo and a system.
Showing 8 of 9
Open weights caught up. A frontier-class model now ships under MIT, and a capable one runs on a single workstation GPU. For clients whose data cannot leave the premises, that changes the answer. This is the anatomy of a build where the network cable is optional.
Requirements documents describe the problem someone already understood. The leverage is almost always in the one they couldn’t, and you only find it from the inside.
A Sydney region alone does not keep your data under Australian law.
Speed means removing the work that was never going to ship.
Tenant isolation is a day-one decision. The database is where it holds.
How we let teams build on real data without ever touching what makes it risky.
The highest-leverage problems never make it into the brief. Here’s why.
A vendor you cannot replace is a dependency, not a supplier. Ownership is the product; who operates the system afterwards is a choice.