An AI engineer interview is not one subject. It moves between machine learning fundamentals, the internals of a transformer, the practicalities of serving a model cheaply, how models are adapted and aligned, how systems are built around them with retrieval and tools, and how any of it is evaluated. Candidates who struggle rarely fail because a topic was too hard. They fail because the ground is wide and their preparation covered a part of it in depth and the rest not at all.
This course is a map and a study path. Each chapter takes one area that interviews return to, explains what is actually being tested, and points you at the lessons on this site that cover it properly. The depth lives in those courses; this one makes sure you know what to read, in what order, and why each thing is asked.
Who it is for
Anyone preparing for a role that builds with language models — AI engineer, GenAI engineer, applied ML engineer, ML engineer on a product team. It assumes you can program and have met machine learning before. It does not assume you have built a transformer or deployed one.
If you are earlier than that, start with Introduction to Artificial Intelligence and come back.
What interviews actually cover
Six areas, roughly in the order a loop tends to move through them.
- Machine learning foundations. Bias and variance, regularisation, the standard algorithms and the trade-offs between them. Asked less than it used to be, and still asked.
- Deep learning and training. Backpropagation, gradient descent and its optimizers, what makes deep networks trainable at all.
- Transformer internals. Tokens, embeddings, attention, positional information, and how generation actually works. The core of a modern GenAI interview.
- Efficiency. Why inference is expensive and the techniques that make it affordable — caching, attention variants, sparsity, low precision, compression.
- Adapting and aligning models. Fine-tuning and parameter-efficient methods, preference training, where training data comes from.
- Building systems. Prompting, retrieval, agents, and the evaluation and failure modes that come with them.
Nobody is examined on all six in one conversation. Every loop touches several.
How to use this course
Read the chapters in order once, quickly. They are short, and the point of the first pass is to see the shape of the subject and find the gaps in what you know.
Then go deep where you are weak, following the links. Each chapter ends with a reading path — the specific lessons that cover its topic, in a sensible order. Those lessons are the real material; this course is the index over them.
Two habits matter more than any single topic.
Be able to say why, not just what. Interviewers rarely stop at a definition. They ask the follow-up — why that design, what breaks without it, what would you give up. Every chapter here flags the follow-up that typically comes.
Recognise a question you have already prepared. The same idea gets asked in a dozen different sentences, and candidates freeze on wording rather than on substance. Every chapter from here on carries a section called How it gets worded — the phrasings the topic actually arrives in. Read those last, after the material, and use them as a self-test: if you cannot answer one aloud, go back to the reading path.
Say what you do not know. The strongest answers are specific about their limits. Guessing confidently is the one failure mode that reads badly in every round, and it is also, as it happens, the failure mode of the models you will be discussing.
The reading path, in one view
| Chapter | Covered properly in |
|---|---|
| Machine learning foundations | Machine Learning Techniques |
| Training neural networks | Deep Learning |
| Transformers and attention | Large Language Models |
| Efficiency and compression | Large Language Models, modules 6–9 |
| Alignment and fine-tuning | Large Language Models, module 10 |
| Prompting, retrieval, agents | Advanced RAG |
| Evaluation | Large Language Models, module 11, and Advanced RAG, module 5 |
| Failure modes and interpretability | Large Language Models, module 12 |
If you prefer history to systems, A Brief History of Deep Learning covers how the field arrived here, which is useful background for the "why is it done this way" questions.
Three to four weeks, part time, is enough to cover this properly if you already know the foundations. One week on transformer internals, one on efficiency and adaptation, one on retrieval and systems, and a few days on evaluation and failure modes. Cramming the transformer in two days does not work — it is the part interviewers probe hardest, because it is the part that cannot be bluffed.