Become an AI Expertv2.1
0 of 0 steps Continue
Stage 4 · Getting inFrom week 14

Module 13: Interviews the 2026 way

Prerequisites and measurable objectives

Prerequisite: Module 12. By the end you can:

  1. Answer the 2026 question bank with numbers from your own projects.
  2. Draw an ML system design for a GenAI app, a recommendation pipeline and an inference service, in under 20 minutes each.
  3. Code attention, a training loop, a tokenizer and the eval metrics from memory while narrating, with or without an assistant on screen.

Primary resource: Alexey Grigorev, AI engineering field guide, interview trends (https://github.com/alexeygrigorev/ai-engineering-field-guide). Finish it before adding anything from the reading list. Hours: 8 guided: 7 on the concept blocks, 1 on the lab and the first pass of the build. Unguided practice on top is yours.

What the rounds look like

A candidate who did 15 AI engineer interviews in six months reports the questions were "why RAG instead of fine-tuning?" and "how did you measure hallucinations?", not math proofs or LeetCode; the best offer came from explaining a 60% cost cut. Another who interviewed at 46 companies saw every format: LeetCode, system design, take-home builds, long LLM discussions. There is no standard loop. Prepare for the software loop plus your own LLM projects.

Concepts, in order

  1. The decisions round (1 hour): for every project, the three hardest decisions, the numbers, the failures you did not fix, and the decision you reversed. Lead with trade-offs.
  2. ML system design (2 hours): a GenAI app end to end (ingestion, retrieval, generation, evals, monitoring, cost); a recommendation or ranking system; an inference service for a 70B model at a latency target and budget. The seniority signal is designing the loop (chunking, context effects, output validation, feedback that improves retrieval), not the prompt. One borderline senior got a soft no for forcing an LLM onto a regression problem.
  3. Coding, with AI in the room (1 hour): "AI allowed, but explain every line" is the dominant take-home format; some companies run one AI-assisted and one AI-banned round; some hand you an unfamiliar repo and grade how you steer the agent, rejecting most candidates; some plant a bad decision to see if you push back. Ask whether AI is allowed. Narrate while you code. Anthropic rebuilt its take-home into puzzles with no debugger that score whether you build your own tools.
  4. The depth probe (1 hour): interviewers pick one resume item and go as deep as they can; a staff MLE rejected candidates who ran out. Know every project to the bottom. Say "I don't know" and reason aloud rather than bluff.
  5. The question bank (1 hour): "Is the model upgrade better, how do you decide before rollout?"; "this RAG answer is confident and wrong, debug it"; "agent or fixed pipeline?"; "estimate per-request cost, then halve it"; "when would you fine-tune?"; "how did you measure hallucinations?" (abstention rate plus a calibrated judge). Red flags: vibes, unbounded loops, cheaper-model-before-measuring, transformer math for an applied role.
  6. Titles and pay (1 hour): level-controlled, the ML premium has plateaued near 1.3x; infra and inference roles sit at the top of the band; "Machine Learning Engineer" titles pay well above "AI Engineer" for similar work, so negotiate the title; every offer in the stories was negotiated and a second offer is the only reliable lever. Decline take-homes that are deployable product work or arrive before any human conversation.

Lab

No lab. Do a timed mock interview with your peer reviewer using the question bank.

Build

Project 14: the application kit. A portfolio page with three projects, each led by a measured result; a resume where every bullet has a number; ten system-design prompts drawn and timed; attention, a GPT block, a training loop with eval, BPE basics, recall@k and pass@k implemented from memory; memory and cost math for 7B and 70B serving; one mock interview recorded.

Deliverables: the kit, the recording, and a one-page "decisions and numbers" sheet per project.

Mastery gate

Your reviewer runs the depth probe on one project for 20 minutes, five levels deep, and you do not run out; the from-memory implementations run.

Quiz

  1. How do you answer "how did you measure hallucinations?" in 60 seconds?
  2. A candidate is asked to debug a confident wrong RAG answer. List the three causes to check first.
  3. What is the seniority signal in LLM system design?
  4. What do you do when a take-home is deployable product work?
  5. Compute serving memory for a 70B model in bf16 at 4k context, batch 8, 8 KV heads, 80 layers, head size 128, and say what hardware it needs.
  6. Why negotiate the title and not just the number?

Primary-source reading

Follow-ups

When you finish this module, paste one of these into your assistant. In Claude Code write /aiexpert-tutor before it; in Codex write $aiexpert-tutor.

  1. Run the depth probe on my Project 9 for 20 minutes and tell me where I ran out.
  2. Ask me the 2026 interview questions from this module and grade each answer for mechanism and measurement.
  3. Give me one from-memory implementation to type with the assistant closed, then review it.

Free to use: code MIT, writing CC BY 4.0, third-party media under its owners’ terms. Licence and terms.