Machine Learning Engineer Machine Learning Interview Questions

For a Machine Learning Engineer the ML round is an implementation round with the mathematics still attached. These 232 questions from Amazon, OpenAI, Snapchat, Pinterest and Apple ask you to write a numerically stable sigmoid and softmax, build scaled dot-product attention out of its pieces, and assign points to their nearest centres without ever materialising the full distance tensor, because the naive version will not fit in memory. Around that sit the questions on why gradients vanish in a deep feedforward network, how optimiser choice and learning-rate schedule interact, what scaling laws predict when you trade model size against training data, and the trade-offs in training a vision-language model or a graph-based recommender. Evaluation gets its own strand: which metric survives class imbalance, and how offline numbers relate to what happens once the model is serving. Amazon contributes 34 and OpenAI 21, a heavier frontier-lab presence than the Data Scientist version of this topic. 61 are rated hard and 131 arrive in a technical screen, so it is often the first real filter. Every one of the 232 has a written solution, and 66% are free to read.