Learning machine learning after years of backend work has been pleasantly familiar in some ways and disorienting in others.
The familiar part is the instinct to make a model of the system before changing it: what goes in, what comes out, where the uncertainty lives, and which measurements matter. The disorienting part is that a perfectly valid program can still produce a deeply unhelpful model. Correctness is not the whole story.
For now, I am treating this as a notebook for the questions that turn out to matter more than the first answers.
A useful starting point
In backend systems, I often begin by looking for the boundary conditions. In ML, the boundary is frequently the data: what it represents, what it leaves out, and whether it resembles the world in which a model will be used.