Model Hosting, governance at scale.

When I arrived, the analytical models teams depended on lived on individual laptops, governed by no one, deployable only by whoever had built them. If that person was on leave, the model was effectively offline. It worked until it very visibly didn't.

So I spent the first few weeks doing almost nothing except listening. The data scientists didn't trust the idea of a central system taking their models out of their hands. The engineers were tired of being the manual bottleneck every time something needed deploying. Both were right, and once I understood what each side was actually protecting, the shape of the platform was obvious. We moved everything onto centralised cloud hosting with stateful APIs and real-time data streams, version control, and self-serve access for the teams consuming the models.

Deployment went from months to days, and for the first time there was real governance: who owns a model, which version is live, who is allowed to consume it. The architecture mattered, but it wasn't the point.