Interview Questions for AI Engineer

Transformer Foundations + "Attention Is All You Need"

1. What is the Transformer architecture?

2. What are the core components of the Transformer model introduced in the paper
"Attention Is All You Need"?

3. What is self-attention, and how does it work?

4. What is multi-head attention and why is it useful?

5. What is positional encoding, and why is it needed?

6. How do encoder and decoder blocks differ structurally in the Transformer?

7. How does the decoder attend to encoder outputs?

8. What are the layers present in a Transformer encoder?

9. What are the layers in the decoder, and how are they different?

10. Why is masked attention used in the decoder?

11. What is the role of residual connections and layer normalization in Transformer
models?

GPT & LLM Architecture

12. What do G, P, and T stand for in GPT?
13. What is a decoder-only architecture? Why is GPT decoder-only?
14. What are the differences between:

◦ Encoder-only models (e.g., BERT)

◦ Decoder-only models (e.g., GPT)

◦ Encoder-Decoder models (e.g., T5, BART)

 15. Can encoder work without decoder? Can decoder work without encoder?
16. Why use decoder-only for text generation, and encoder for classi cation?
17. Is it possible to ne-tune just the decoder part of a model? What are the trade-
offs?

18. Can you mix and match encoder-decoder con gurations across use cases?
19. What are causal vs bidirectional attention types?

Generative AI Concepts
20. What is Generative AI? How is it different from traditional discriminative ML?

21. Explain how text generation works in LLMs like GPT.

22. What are prompt engineering and prompt tuning?

23. What is ne-tuning an LLM? When would you do it?

24. What is LoRA (Low-Rank Adaptation), and why is it ef cient for ne-tuning?

25. What are the trade-offs between RAG and LLM ne-tuning?

26. When would you use OpenAI models (GPT-4) instead of open-source models?

NLP Basics + Vectorization

27. What is tokenization in NLP?

28. What is the difference between tokenization and chunking?

29. What is a chunk size and why is 100 often chosen?

30. What does chunk overlap mean (e.g., overlap = 50)?

31. What are dynamic chunking methods, and when are they useful?

32. What is BoW (Bag of Words) and its limitation?

33. What is TF-IDF and how is it calculated?

34. What is the difference between TF-IDF and embeddings?

35. What is vectorization in NLP and what are different methods for it?
fi
fi
fi
fi
fi
fi
fi
 36. What are embedding models? Give examples of embedding models you've
used.

37. Can we use K-means clustering instead of similarity search? Why or why not?

Embeddings & Vector Databases

38. What are text embeddings?

39. How are embeddings used in a RAG system?

40. What is a vector database?

41. Which vector database did you use (e.g., Pinecone, FAISS, Weaviate), and why?

42. How do you store and retrieve data using a vector DB?

43. What are trade-offs between Pinecone and FAISS?

44. How does the embedding dimensionality affect retrieval performance?

Retrieval-Augmented Generation (RAG)

45. What is RAG, and why do we need it?

46. Explain the architecture and work ow of RAG.

47. What are the major components of a RAG pipeline?

48. How does the retrieval process work in RAG?

49. What is semantic search and how does it work?

50. What is hybrid search (semantic + keyword)?

51. What is segment-level search or hierarchical retrieval?

52. What kind of embedding models can be used in RAG?

53. Why did you choose RAG over ne-tuning?

54. What is the difference between Traditional RAG, Agentic RAG, and Graph RAG?

55. What is Graph RAG and when should it be used?

56. What is the difference between Dense RAG and Hybrid RAG?
fi
fl
RAG Evaluation
57. How do you evaluate a RAG system?

58. What are key RAG evaluation metrics?

• Relevance / Precision

• Faithfulness / Groundedness

• Latency

• Hallucination Rate

• Source Attribution

59. How do you perform human vs automated evaluation?

60. What tools exist for RAG evaluation (e.g., RAGAS, TruLens)?

Agentic AI & GenAI Agents

61. What is Agentic AI?

62. What are the core components of an agent (planner, memory, tool, executor)?

63. What are different types of agents (tool-using, reactive, planning, long-term
memory)?

64. How do agentic work ows improve over static chains?

65. What is the difference between a LangChain chain, a graph, and an agent?

66. What is Agentic RAG, and how does it enable adaptive retrieval?

67. What is Agentic Graph RAG?

68. What are the advantages and challenges of Agentic systems?

GenAI Agent Frameworks

69. What is LangChain? What are its core components (chains, tools, agents)?

70. What is LangGraph and how does it extend LangChain?

fl
 71. What is the advantage of LangGraph for building work ows?

72. What is AutoGen? When should it be used?

73. What is CrewAI, and how does it manage multi-agent collaboration?

74. Compare LangChain, LangGraph, AutoGen, and CrewAI – when to use what?

Coding Questions (Basic to Mid-Level)

75. Write a function to nd the second-largest number in a list.

76. Reverse a string without using Python slicing.

77. Remove duplicates from a list while preserving order.

78. Count frequency of each word in a paragraph.

79. Implement K-means clustering from scratch.

80. Check if a string is a palindrome.

81. Convert a list of strings to a TF-IDF matrix using sklearn.

82. Write code to chunk a document with overlap size 50 tokens.

Classical Machine Learning + Managerial Round

83. What are the key assumptions of linear regression?

84. In a classi cation problem, why is recall more important than precision?

85. How would you x a high variance (over tting) model?

86. Your churn prediction model has low recall. What steps will you take?

87. Explain the use of Random Forest to a non-technical business user.

88. Which metrics would you use for a regression problem predicting house prices?

89. Your model works well on test data but fails in production — why?

90. How do you detect and mitigate data drift or concept drift?

91. What is bagging and boosting?

92. What is bias and variance tradeoff?

fi
fi
fi
fi
fl
 93. What is confusion matrix? How and why we use it?

94. What steps will you take to ensure a model works well on real-world data?

95. If you are working on a Retail project and have million of rows and 100+ features
how you will approach a project. Consider the data is take from a source directly
without any modi cations of cleaning. Explain each steps in detail and what
approach will you take to train your model and which model will you use? (Key:
Keep asking multiple followup and cleari cation questions to solve this question.
Usually asked in 2nd or 3rd round of interview.)

MLOps + End-to-End GenAI Project Deployment

96. Describe your end-to-end GenAI pipeline.

97. Why did you use FastAPI for deployment?

98. What is the role of AWS S3 in your pipeline?

99. What is EC2, and how do you use it for model hosting?

100. How do you handle model monitoring, versioning, rollback?

101. How does your pipeline support real-time inference vs batch inference?

Final Note:
Make sure you go through all your resume projects in detail.
Most of the interview questions are going to be asked from your own end-to-end
projects, especially if they include LLMs, GenAI, RAG, or deployment.

Many of the questions are jus a followup questions from your response. So make sure
you reply smartly and use only those terms which you are con dent about.

Much Love, Happy Learning

ALL THE BEST !!!
fi
fi
fi