Big Tech platforms AI Engineer Interview Loop

Unlike every other role here, this screen doesn't start with SQL. It's ML and generative AI basics, asked fast. Why is attention expensive at long context? What does the KV cache do at inference? How do encoder and decoder models differ? Which decoding strategies exist, and what is retrieval-augmented generation for? There's a coding problem too. The bar is fluency: short, correct answers that explain how it works.

The coding round at medium to hard, plus the SQL and Python data handling that AI systems are built on. These are the same order, event and user tables every other role queries, because the features have to come from somewhere.

LLM and AI product design: a RAG system, a multi-step LLM agent, search-intent prediction, an on-device language model, and RAG-powered slide generation. Scope the task and choose between prompting, retrieval and fine-tuning. Design the evaluation, then talk about latency, cost, hallucination and jailbreak defenses.

RAG versus fine-tuning, and how to evaluate RAG systems and generative models. LLM hallucination and its biases come up too, and A/B testing with network effects when you compare two model versions. So does the classical ML that still sits under most production systems.