Picture a support-ticket assistant: it reads a customer's ticket thread and account notes, then drafts a reply for a human agent to send. Early in an AI (artificial intelligence) engineering interview about a feature like this, the interviewer asks, "Which model would you use, and what will limit it?"

A good answer does not need transformer math. It needs a working picture of what a large language model (LLM) does when you call it: it reads your text as tokens, fits them into a fixed window, and writes the reply one sampled token at a time. Almost every limit you will hit in production, from cost to slow replies to confidently wrong answers, follows from that loop.

That loop is also why an LLM fits open-ended text work, such as drafting this reply or tagging a free-form ticket, and is the wrong tool when a rule, a regex, or a database lookup already gives the exact answer, or when the output must be exact arithmetic.

The question itself usually gets a few minutes, but everything said later in the round assumes this vocabulary, so a shaky answer here colors the rest. A mid-level candidate is expected to use the terms correctly and pick a sensible model with one reason. A senior candidate predicts which limit the feature hits first, explains why output varies even when it "shouldn't", and says what evidence would make them switch models.

This tutorial covers the engineer's view of the model: tokens and the context window, how input and output length drive latency, how the next token is sampled, why models hallucinate, and how those behaviors turn into a model choice.