.png){kind=link}
Machine learning covers a lot of ground, from a small classifier that routes support tickets to a fine tuned language model that produces output in a particular house style. The common thread is a system that learns behaviour from examples rather than being told the rules directly.
What the work involves
Framing the problem. This is the step that decides whether anything else can succeed. What exactly is being predicted, from what information, and what does a correct answer look like. A large share of projects that fail were unanswerable as posed, and no amount of modelling recovers from that.
Understanding the data before modelling it. Where it came from, what it misses, how it is biased by how it was collected, and whether the labels mean what everyone assumes. Most of the effort in a real project lands here.
Choosing the smallest approach that works. A logistic regression that is understood and maintainable often beats an elaborate model nobody can debug. Complexity is a cost paid every time the system needs to change.
Honest measurement. Accuracy on a test set that leaks information from the training set is not a result. Choosing a metric that reflects the actual cost of each kind of mistake matters more than the headline number.
When not to use it
Worth saying plainly, because it saves people a great deal of money.
If the rules are known and stable, write the rules. If there are not enough examples, there is nothing to learn from. If a wrong answer is unacceptable and there is no way to review the output, a probabilistic system is the wrong shape of solution.
Machine learning is a good fit when the pattern is real but hard to articulate, there is enough data to show it, and being right most of the time is genuinely useful.
How it connects
In practice most of the machine learning inside a modern product is not a model trained from scratch. It is a pre-trained model used well, with the engineering effort going into the data around it and the measurement of whether it is working.
That is why this sits alongside AI engineering rather than apart from it.