Most ML courses teach modelling and call it engineering. The six production skills a real ML engineering course must cover, the warning signs to watch for, and how we structure ours.
"Machine learning engineering" has become the label for two different jobs. One is building models. The other is making models survive contact with real data, real users and real infrastructure. Most courses teach the first and call it the second, which is why so many graduates can train a classifier in a notebook and cannot get it in front of a customer.
If you are comparing ML engineering courses, here is what the production half of the craft actually contains, and how to check whether a course covers it.
Every model is downstream of a pipeline, and most production failures are pipeline failures: a schema changes upstream, a null appears where it never did, a feature is computed differently at training time and serving time. A serious course spends real time on orchestration, validation and the warehouse the features come from. If the course starts at "load the CSV", it is a modelling course.
Accuracy on a held-out set tells you almost nothing about whether a model is safe to ship. You need offline metrics that mirror the decision the model drives, a baseline you must beat, slice-level analysis so you know who it fails for, and for LLM systems, an evaluation set and a judge you trust. Ask whether the course makes you build an eval harness. If not, you will build your first one on a live system.
A model that answers in four seconds when the product needs 300 milliseconds is not a model, it is a demo. Batch versus real-time serving, caching, model size trade-offs and the actual bill for inference are engineering decisions, and they belong in the curriculum.
Models degrade quietly. The course should cover what to log, how to detect data drift and prediction drift, and how to decide when to retrain. This is the part that keeps the system working at three in the morning when nobody is watching.
For LLM applications in particular: retrieval quality, prompt injection, hallucination handling, rate limiting and what the system does when the model is confidently wrong. A course that treats "add a system prompt" as the safety chapter is not preparing you for an enterprise review.
The only reliable test of whether someone can do production ML is whether they have done it. The assessment should be a system you built, running, that you explain and defend to people who ask hard questions. Multiple-choice quizzes measure recall, not engineering.
"Covers 40+ algorithms." Breadth is a modelling-course signal. Production teams use a handful of model families and spend their time on everything around them.
No named instructor. Production judgment is tacit. It transfers from a person who has been on call for a model, not from a slide deck.
Self-paced with no deadline. Production ML is a team sport with deadlines. A course with neither rarely gets finished, and never gets defended.
No mention of cost. If the word "budget" does not appear in the syllabus, the course has not met a real product owner.
We split the discipline into the courses that match how teams actually hire.
AI & Data Science is the "why it works" course: twelve weeks from statistics to machine learning and LLMs, concept before code, for analysts and the technically curious. $4,620 incl. GST.
AI Engineering & Production is the production half: twelve weeks live, capped at twelve, covering pipelines, serving, evals, guardrails, monitoring and agentic coding, finishing with a system you ship and defend. $4,950 incl. GST.
Data Engineering builds the platform underneath both: pipelines, warehouses and orchestration, twelve weeks, ending with a production data platform. $4,350 incl. GST.
Generative AI Systems & Safety goes deep on RAG, evaluation, guardrails and red teaming for LLM systems you have to defend in front of a security or risk team. $5,450 incl. GST.
For someone changing careers, the Applied AI Capstone runs the full arc over twenty-four weeks, entry by assessment.
Every one of these is live, senior-led and assessed on a real artefact. The full list, with intake dates where they exist, is on the programs page. If you are not sure whether you are ready for the production course or should start with the foundations, the "ready if / not yet if" section on each page is written to answer that, and a qualification call settles it in thirty minutes.
