ML platform

The practical argument for hosting this layer is that laptops sleep. A six-hour job, a labelling project running across a team, an experiment history that has to outlive somebody’s machine — all of it wants a server, and none of it wants a per-seat SaaS contract.

There is a data argument as well. Training sets and annotation queues are often the most sensitive material in an ML project, more so than the model itself. Keeping them inside your own boundary is the simplest version of that problem solved.