#Interview question bank

Every question from across the handbook, consolidated. Each links back to the page that answers it properly.

⭐ marks the questions that most reliably separate candidates — usually because the obvious answer is wrong, or because most people stop one level short.

#How to drill this

Do not read the answers. Cover them, answer out loud, then check.

The failure mode of a question bank is recognition without recall: you read the answer, it feels familiar, you conclude you know it, and then you cannot produce it under pressure. Say it out loud — the gap between "I know this" and "I can say this in sixty seconds" is where interviews are lost.

Time-box to two minutes per question. Most answers should take sixty to ninety seconds. If you cannot finish in two minutes you are including material that is not earning its place.

#Round 1 · Foundations

Asked to check you understand the machine, not just the API.

#Round 2 · RAG and retrieval

The largest section, because it is the most-asked area.

#Round 3 · Evaluation

Where most candidates are weakest, and where a good answer is most noticeable.

#Round 4 · Systems, serving and optimization

#Round 5 · Agents, safety and adaptation

#Round 6 · Judgement, product and business

Asked at senior level, and where technical candidates most often stumble.

#The twelve that matter most

If you have one evening, drill these. They cover the widest ground and each has a counter-intuitive core that most candidates miss.

1. Why does quantization make inference faster? — bandwidth, not arithmetic
2. The score went up 4 points on 50 items. Ship it? — no; ±11 point interval
3. How do you know your judge is any good? — kappa against human labels
4. How do you prevent prompt injection? — you don't; you bound the blast radius
5. Why do long agent chains fail? — reliability compounds multiplicatively
6. RAG or fine-tuning? — knowledge versus behaviour
7. RAG quality is poor. Where first? — the retrieval ceiling
8. Is chain-of-thought an explanation? — no, and saying why is the signal
9. You pruned to 90% and it's not faster. — unstructured sparsity does nothing
10. Can you make it fair? — not in all senses at once; it's a proven result
11. What's our moat? — not the prompt, model or pipeline
12. When should we not use an LLM? — when a deterministic solution exists

#Answering well, mechanically

Lead with the answer, then justify. "No — at n=50 the interval is ±11 points" beats three sentences of preamble arriving at the same place.

Name the trade-off. Almost every question here has one. Naming it unprompted is most of what "senior" sounds like.

Say what you would measure. The strongest ending to any answer is how you would know you were right.

Admit the limit. "I would need to check the current pricing" is stronger than a confident wrong number. Interviewers are calibrating whether they can trust what you say — a single fabricated detail costs more than the answer gains.

Use your own work. You have an eval harness with a golden set, a gate, and a failure-injection test you watched go red. Nearly every question in Rounds 3 and 4 can be answered with "here is what I actually did" — which beats any theoretical answer available to anyone else in the pipeline.