aiengineering.guideaiengineering.guide

ROADMAP · ~14 hours over 2 weekends

Interview cram · 2 weeks out

You have an AI-eng interview in the next 1-3 weeks.

Start: What is an LLM, really?11 stops · ~14 hours over 2 weekends

Who this is for

You have an AI-engineering loop in the next 1-3 weeks. You’ve built with LLMs but haven’t formalized what you know into interview-shape answers.

This isn’t a substitute for engineering practice — if the roadmap here reads as unfamiliar territory rather than review, defer the loop, use /roadmap/swe-to-ai for 3 months, then come back.

What to actually do

  • Weekend 1: Read the 6 interview cards. Answer each out loud (record if you can). If any takes > 3 minutes to explain, read the linked lesson.
  • Weekday evenings: One lesson per evening.
  • Weekend 2: Whiteboard the system-design question. Read the LLM Gateway lesson. Whiteboard the LLM Gateway design.
  • Day before: Re-read the case file. Don’t read anything else.

Company tag hint

If you know your interviewer’s employer, filter /interview by company. Most companies have a repeated question or two.

THE PATH

11 stops, in order

Phase 1Warm up1 stop

  1. 1

    Lesson

    What is an LLM, really?

    10-minute refresh. Skip if you can explain temperature vs. top-p already.

Phase 2Rapid-fire cards5 stops

  1. 2

    Interview card

    What is a token, and why does the same prompt cost different amounts on different models?

    Come in able to explain this in 15 seconds without pausing.

  2. 3

    Interview card

    Temperature and top-p both control randomness. What's the difference, and when would you reach for each?

    Same. If both of these throw you, spend a day on Foundations first.

  3. 4

    Interview card

    Give me three cases where semantic search over embeddings retrieves the wrong document — and explain why.

    The RAG failure-modes question comes up in ~70% of loops now.

  4. 5

    Interview card

    What's the difference between an eval and a benchmark, and why would a team run both?

    If you say "we tested on MMLU" in an interview and mean "eval" you'll lose points.

  5. 6

    Interview card

    How would you defend a customer-support agent against prompt injection when it can send emails and read a customer database?

    The security round question. Own the "architectural, not regex" answer.

Phase 3System design1 stop

  1. 7

Phase 4Whiteboard-ready depth4 stops

  1. 8

    Lesson

    Tool use, from three lines of Python

    Every agent-shaped question depends on this.

  2. 9

    Lesson

    Evaluating a RAG system without lying to yourself

    You will be asked "how would you know your RAG works." Have an answer.

  3. 10

    Lesson

    Design an LLM gateway

    The FLAGSHIP. Read once, then whiteboard it three times.

  4. 11

    Case file

    The eval pipeline that lied for a month

    Read the failure-mode retrospective. Interview gold.

NOT COVERED HERE

What this roadmap skips

  • The full roadmap. If you have 6 months, use `/roadmap/swe-to-ai`.
  • Deep math. If they're asking you calculus on an AI-eng loop, it's not an AI-eng role.
  • Prompt engineering as a technique. Interview questions are about principles, not tricks.

← All roadmaps

Enter to go · Esc to close