Interview Prep Hub

AI Engineer Interview Prep

Role-neutral preparation for AI engineer loops. The sixteen questions that keep appearing in 2026 interviews, each worked as a scenario with the probes that follow it: RAG, agents, evals, LLM system design, and production debugging, all on one running example so the answers connect.

Sixteen questions, one companythe questionscopemechanismtrade-offmeasurewhatbreaksfirstHarbourline InsuranceRAG · agentsevals · voicebatch · uploads40k chats/day6k claims/dayanswer aloud first · then open the probes
If you only have an hour

Read 00 Start Here for the running example and how the rounds are scored. Then do 01, 04, 07 and 16: between them they cover what every loop asks, and each ends with questions you should answer aloud.

Study pathsPick the plan that fits the time you have
7+ days

Day 1: 00, then 01–03 (retrieval). Day 2: 04–06 (agents). Day 3: 07–08 (evals). Day 4: 09–10. Day 5: 11–12. Day 6: 13–15. Day 7: 16, then redo every Check yourself aloud.

2–3 days

00, 01, 03, 04, 05, 07, 08, 10, 13, 15, 16. Skim the rest by their Takeaway sections.

< 24 hours

00, 01, 04, 07, 16. Then read only the Takeaway of 10, 13 and 15.

Section A · Orient
00 Start Here How AI engineer loops are scored, the running example every chapter uses, and how to work through the sixteen questions.
Section B · RAG and retrieval
01 The Wrong Answer When the Right Document Exists A RAG answer is wrong even though the correct document is in the index. The six places it can fail, the trace that tells you which one, and the fix for each. 02 Chunk Size and Overlap How to choose chunking for a corpus you have never seen: what the embedding model, the questions and the generator each constrain, and how to measure the choice instead of guessing. 03 Hybrid Search and Reranking When dense retrieval alone misses, when BM25 and fusion fix it, when a reranker earns its latency, and how to say which one your data needs.
Section C · Agents
04 Agent or Workflow The decision interviewers most want to hear you make well: when a model-driven loop beats a deterministic pipeline, when it does not, and the hybrid most production systems end up with. 05 A Claims Agent Under a Token Budget Design an insurance-claims agent that uses RAG and keeps token cost under control: context budgeting, caching, routing, and the metrics that prove it. 06 A Customer-Service Agent The full design conversation: scope, tools, escalation, memory, safety, evals and rollout, with the trade-offs an interviewer will push on.
Section D · Evals
07 The Eval Set and the Ship Decision How to build an eval set that predicts production, how big it needs to be, how to grade it, and how to turn a score into a yes or no on shipping a change. 08 Evaluating an Agent Beyond Its Final Answer Trajectories, tool calls, cost and steps, partial credit and safety: how to judge an agent that reached the right answer the wrong way, or the wrong answer for a good reason.
Section E · LLM system design
09 A Batching System for LLM Queries Design end-to-end batching for LLM work: the request path, queues and windows, continuous batching at the server, provider batch APIs, and the latency the batching costs. 10 Latency and Cost at Scale The levers in order of payoff: prompt and prefix caching, model routing, output limits, speculative decoding, streaming and parallel tool calls, with the numbers to quote. 11 A Large-File Uploader for a Multimodal Chat Design the path from a 2 GB video or a 300-page PDF on a phone to a model that can answer questions about it: resumable upload, processing, storage, and what the chat sees while it waits. 12 A Real-Time Voice Application Design voice in and voice out with a model in the middle: the latency budget, streaming STT and TTS versus speech-to-speech models, turn-taking, interruption and tools.
Section F · Production and debugging
13 An LLM Behind a Reliable Contract How to put a probabilistic model behind an API your callers can depend on: schemas, validation, retries, timeouts, fallbacks, idempotency and versioning. 14 The Agent That Loops Forever A live-debugging round: an agent that never terminates on some inputs. How to find the cause from traces and how to make the loop impossible, not just unlikely. 15 Confident and Wrong in a High-Risk Workflow A model gives fluent, wrong answers where the cost of error is high. What to change in retrieval, prompting, abstention, verification and the human loop, and in what order. 16 The System You Shipped The behavioural spine of every loop: how to walk through an AI system you built, its worst production failure, and how you measured the fix, in a form interviewers can score.