100 AI Engineer interview questions worth preparing

AI Engineer interviews cluster into a few areas, and depth in one won’t save you if another is blank. Here’s the full bank, organised by category, so you know exactly what to prepare. Don’t memorise answers — understand each concept well enough to explain it simply and tie it to a project you actually built. That’s what gets the offer.

Python & programming fundamentals

1. What makes Python the default language for AI/ML work rather than the alternatives?
2. When would you reach for a list vs a tuple vs a set vs a dictionary, and what’s the trade-off?
3. What are the lookup, insert, and delete costs of Python’s core data structures?
4. How do generators differ from iterators, and why are generators easier on memory?
5. What problem do decorators actually solve? Walk through a real use case.
6. What do *args and **kwargs let you do in a function signature?
7. What’s the practical difference between a deep copy and a shallow copy?
8. Multithreading or multiprocessing — how do you decide which a workload needs?
9. What is the Global Interpreter Lock, and how does it affect performance?
10. Explain the core OOP concepts using an AI application as the example.

Math, probability & statistics

11. What is the bias-variance trade-off, and how does it show up in a real model?
12. Walk through Bayes’ theorem using a machine-learning scenario.
13. How is likelihood different from probability?
14. Why does the Central Limit Theorem matter in practice?
15. What separates a Type I error from a Type II error?
16. When would you optimise for precision, recall, or F1 — and why?
17. Why can accuracy be a misleading metric on an imbalanced dataset?
18. How do you explain the difference between correlation and causation?
19. What does a p-value actually tell you, and what does it not tell you?
20. Why do we scale or standardise features before training, and when is it required?

Classical machine learning

21. Compare supervised, unsupervised, and reinforcement learning with examples.
22. How does linear regression work under the hood?
23. Why do we use logistic regression for classification instead of linear regression?
24. What breaks when you force linear regression onto a classification task?
25. How does a decision tree decide where to split?
26. What does a random forest add over a single decision tree?
27. Bagging vs boosting — what’s the core difference in approach?
28. Why has XGBoost become a default choice for tabular data?
29. Explain how K-Means groups data points.
30. How do you decide on the right number of clusters (K)?
31. What is PCA for, and when should you actually apply it?
32. What signals tell you a model is overfitting?
33. When would you choose L1 regularisation over L2, and vice versa?
34. Why is cross-validation more reliable than a single train-test split?
35. What strategies do you use to tune hyperparameters?

Deep learning

36. In plain terms, how does a neural network actually learn?
37. Explain backpropagation without the heavy math.
38. How does gradient descent adjust the weights step by step?
39. What causes vanishing and exploding gradients, and how do you fix each?
40. Compare sigmoid, tanh, ReLU, and GELU — when does each make sense?
41. How does batch normalisation differ from layer normalisation?
42. Why does dropout help a network generalise?
43. When would you pick a CNN vs an RNN vs a Transformer?
44. What problem were LSTMs designed to solve?
45. Which weaknesses of RNNs did Transformers overcome?
46. Walk through the Transformer architecture at a high level.
47. What is self-attention doing, intuitively?
48. Why use multiple attention heads instead of one?
49. Why do Transformers need positional encoding, and what is RoPE?
50. How do GPT, BERT, and T5 differ in design and in what they’re used for?

LLM fundamentals

51. In your own words, what is a large language model?
52. What are tokens, and why do they matter for cost and context limits?
53. What are embeddings, and what do they capture?
54. How is cosine similarity used when working with embeddings?
55. What is a context window, and why is its size a constraint?
56. Explain the “lost in the middle” problem in long contexts.
57. Why do LLMs hallucinate in the first place?
58. What are practical ways to cut down on hallucinations?
59. Compare zero-shot, few-shot, and chain-of-thought prompting.
60. When do you prompt-engineer a model vs fine-tune it?
61. How do SFT, RLHF, and DPO differ as fine-tuning approaches?
62. What is LoRA, and what problem does it solve?
63. How does QLoRA build on top of LoRA?
64. What is model quantization, and what’s the trade-off?
65. What is the KV cache, and why does it speed up inference?
66. How do temperature, top-p, and top-k each shape the output?
67. What’s the difference between a presence penalty and a frequency penalty?
68. What is a Mixture of Experts model, and why would you use one?
69. Why is inference a different engineering problem from training?
70. How would you bring down LLM latency and cost in production?

RAG (Retrieval-Augmented Generation)

71. What is RAG, and what problem does it solve?
72. When is RAG the better choice over fine-tuning?
73. What are vector embeddings in the context of retrieval?
74. How do you decide chunk size and overlap when splitting documents?
75. What is a vector database, and how does it search?
76. Compare FAISS, Pinecone, Chroma, and pgvector.
77. Dense retrieval, BM25, and hybrid search — when do you use each?
78. What is reranking, and why do cross-encoders help?
79. Explain query rewriting, HyDE, and multi-query retrieval.
80. How do you actually measure whether a RAG pipeline is any good?
81. What are the common ways a RAG system fails?
82. Walk through how you’d build a production-grade RAG chatbot.

Prompt engineering & LLM systems

83. How do prompt engineering, RAG, and fine-tuning compare for the same problem?
84. What’s the role of a system prompt vs a user prompt?
85. How do you get reliable structured output (like JSON) out of an LLM?
86. What is function/tool calling, and when do you use it?
87. What is a prompt injection attack, and how do you defend against it?
88. What does good context engineering look like in practice?
89. How do you reduce token usage without hurting answer quality?
90. How do you pick the right model for a given task?

AI agents & MCP

91. What actually makes something an “AI agent” rather than a script?
92. Where’s the line between a workflow and a true agent?
93. Explain the ReAct pattern for agents.
94. How do short-term and long-term memory differ in agent systems?
95. Compare LangGraph, CrewAI, and AutoGen.
96. What is MCP (Model Context Protocol), and what does it standardise?
97. When does a multi-agent setup make sense over a single agent?

Production AI & evaluation

98. How do you evaluate an LLM application once it’s live in production?
99. How do you monitor and keep improving an AI system after deployment?
100. Design an end-to-end AI application (LLM + RAG + agents + evaluation) for a real use case, start to finish.

How to use this: don’t try to cram all 100 the night before. Work through one category at a time, and for each question, practise saying the answer out loud in two minutes and tying it to something you’ve built. The ones you stumble on are your study list. Depth plus a real project beats memorised definitions every time.