How ML Runs in Production

One sentence from Chapter 7 deserves its own closing spotlight, because it quietly decides what many readers should study next. Here it is again: a deployed model is a program in production. Plateful's delivery model does not float in some mathematical realm — it lives on servers, ships in a container like any other software, redeploys through a pipeline whenever retraining produces a new file, and when it breaks at 3 a.m., it pages a human.

Read that list again and notice what's missing: nothing in it is specific to machine learning. Servers, containers, pipelines, monitoring — that is the ordinary software world, and the model is cargo riding on it. The ML world runs on the ordinary software world, and that is good news, because the ordinary software world is teachable — this catalog teaches it. This page is the map of that stack, with each layer's course named.

The Stack Under the Model

Start at the top and look down. The model-as-a-service from Topic 33 — a program that answers prediction requests — has to run somewhere, and that somewhere is almost always the cloud: computers you rent by the hour in someone else's data center instead of buying and racking your own. The model file is packaged and versioned like any software, loaded onto those rented machines, and pointed at by the app.

Cloud from Zero is the from-scratch door into that layer: what the cloud actually is, how the big providers compare, what regions and billing mean — no prerequisites, same register as this book. Everything Plateful's models run on, that course explains from the ground up.

Shipping and Automating

Now recall what Topic 34 established: models age, so retraining is scheduled work — which means the shipping never stops. Every retrain produces a new file that must be tested, versioned, and rolled out without breaking Friday dinner rush, and no team does that by hand for long. They automate it: pipelines that carry a change from "ready" to "live" through checks, the same way factories replaced hand assembly.

That discipline has its own name and its own course: DevOps for Beginners — how modern software gets built, tested, and shipped automatically rather than by hand and hope. (When you're ready for the tooling under it, the catalog's Git course is the standard next step.) None of it was invented for ML; all of it is what keeps ML running.

The Specialization with a Name

Running models in production has grown enough of its own toolbelt to earn a title, and you'll meet it in job ads: MLOps — DevOps thinking applied to the model lifecycle you learned in Chapter 7. That one sentence is genuinely the whole definition: the retraining schedules, the version rollouts, the monitoring for silent aging, treated as an engineering discipline rather than a heroic effort.

It is also a real career direction, and it sits exactly at this catalog's intersection: half of it is this book's Chapter 7, and the other half is the cloud-and-pipelines world the figure above just mapped. People who hold both halves are chronically scarce.

Choosing Your Door

So which door is yours? Honest test: which part of this book made you lean in? If it was the machine room — where the model physically runs, how the file gets from a training run to a million phones — start with Cloud from Zero, or go one floor further down with Computing Foundations from Zero, which explains what a computer, a program, and a network actually are, from nothing. If it was the models themselves — the framing, the training, the honest evaluation — the next page lays out that path.

And note that this is a fork in your reading order, not in your future: the two routes reconverge in every real ML team. Topic 50's brigade needs both wings — the people who make models worth shipping, and the people who keep shipped models alive.