AI Eng · Interview Lab — Course Map
AI Engineering and GenAI with JS teach you to build LLM apps. This section is the practice for being interviewed on them.
Read this first. It takes five minutes.
What AI engineering interviews actually test
Interviews for AI product and AI engineer roles have settled into a few recognisable rounds. None of them is "explain what a transformer is":
| Round | What they hand you | What they're scoring |
|---|---|---|
| LLM system design | "Design a support assistant over our docs" / "an agent that can issue refunds" | Retrieval, model choice, tools, guardrails, evals, cost and latency, and what happens when the model is wrong |
| Code review of an LLM feature | A repo with a chatbot, a RAG pipeline or an agent | Whether you see prompt injection, unsafe output handling, runaway costs, and the ordinary bugs around them |
| Your own project | "Walk me through the AI system you built" | Whether you measured it, what broke, what you'd change |
| Practical / take-home | Build a small feature with an LLM API | Clean API use, structured output, streaming, error handling, and an eval that shows it works |
The first two need practice that reading doesn't give you. That's this section.
How LLM bugs differ from ordinary bugs
Everything in Bug Hunt still applies to LLM apps. On top of that, LLM apps have one new property: part of the program is a component that reads instructions from its input. That creates new failure modes:
- Everything that reaches the prompt can steer the model, not just the user's message: a retrieved document, an email, a web page, a tool's result.
- Everything the model produces is untrusted input to your code: text, JSON, tool arguments, URLs, SQL.
- The model is probabilistic and expensive. It can stop early, refuse, loop, or cost real money per call, so ordinary reliability bugs (retries, timeouts, caching) have new consequences.
Reading order
| # | Chapter | You'll be able to… |
|---|---|---|
| 1 | The LLM Bug Catalogue | Name the bugs only LLM apps have, graded L1–L6 and mapped to the OWASP Top 10 for LLM Applications |
| 2 | LLM Bug Hunt Drills | Find them in code: prompts, parsing, output handling, agents, RAG, streaming, cost, evals, plus a mock review |
| 3 | LLM System Design: The Method | Answer any "design an LLM feature" question in a structure interviewers recognise |
| 4 | LLM System Design Drills | Practise five design prompts with model answers |
How this connects to the rest of the site
- AI Engineering teaches the concepts. Chapters on evals, security and cost and latency are the deep versions of what these drills test.
- GenAI with JS builds the features: the agent loop, RAG, streaming, durable workflows.
- Bug Hunt is the method and the general catalogue. The drills here use the same click-to-spot format and the same L1–L6 levels.