Rehearsing AI-eng interviews out loud (SoloMock)
In one line: The AI system-design page and the defend-your-portfolio drill tell you what the rounds look like — this page is about rehearsing them out loud, solo, with each round mapped to a specific mock problem so your reps are deliberate.
By 2026 the AI-engineer loop is its own thing. You'll get a RAG design round, almost certainly an evals round (interviewers call evals "the new system design"), an agent / tool-use round, and — increasingly — a prompt-injection / agent-security round. You can read about all four and still fumble when someone says "ok, design the eval for this agent, out loud, now." The fix is reps where you talk through the tradeoffs under time pressure, not just nod along to a blog post.
The tool: SoloMock
SoloMock is a free verbal mock-interview app (a companion project to this guide). You talk to an AI interviewer over voice while you sketch or code in a real editor. Its AI Engineer track doesn't grade Big-O — it pushes on the things these interviews actually probe: tradeoffs, failure modes, cost/latency math, and evaluation rigor. Pick the AI Engineer track to filter to the problems below.
The 2026 AI-eng loop, and what to rehearse for each round
| Round | What it actually tests (2026) | Rehearse with |
|---|---|---|
| RAG system design | End-to-end: ingestion, chunking, retrieval, citations, escalation — and the cost/latency math | Design a RAG support chatbot · Chunking strategy |
| Evals ("the new system design") | Build a gold set, LLM-as-judge, catch regressions before prod — the most heavily weighted skill | Build an eval for a multi-tool agent · Faithfulness eval |
| Agents / tool use | An agent loop with tools, memory, termination, and the failure modes | Implement a ReAct agent loop |
| LLM / agent security | Prompt injection on a tool-using agent, the lethal trifecta, defense in depth | Defend an agent against prompt injection |
| Prompt eng / structured output | Robust extraction with validation + retries; structured-output mode | Structured extraction with retries |
Candidates pour prep into RAG diagrams and skip evals — then lose the loop on "how would you know this agent works?" In 2026 write-ups, evals engineering (golden sets, LLM-as-judge, regression gates) is repeatedly called the most heavily weighted and the most common failure point. The non-obvious depth: a right answer reached via a broken or wasteful tool path is still a regression, so you evaluate the trajectory, not just the final answer — and you must keep the judge honest (temperature 0, neutral prompt, calibrated against human labels). Drill the agent-eval problem until you can design that out loud.
How to actually practice
- One round, out loud. Start the timer, ask clarifying questions first (latency budget, eval criteria, data freshness — AI systems live or die on these), sketch the architecture verbally before any code.
- Force yourself to name failure modes. Hallucination, prompt injection, retrieval misses, cost blowups, malformed JSON. The AI interviewer won't volunteer them — enumerating them unprompted is the senior signal.
- Always close the loop on evaluation. For any system you design, answer "how would you measure that it works?" without being asked. That habit alone separates strong from average.
- Then go deeper. Pair this with the AI system-design round (the canonical "Design ChatGPT / Cursor / Perplexity" questions) and the defend-your-portfolio drill.
- All architecture, no evals. If you can draw the RAG pipeline but can't design its eval, you'll lose the loop. Practice the eval rounds as hard as the design rounds.
- Treating prompt injection as "add a system-prompt line." The interviewer wants to hear that the model's text output must never be the security boundary — dangerous actions get gated by deterministic policy code. "Just tell it not to obey" is the wrong answer.
- Hand-waving cost and latency. "It'll be fast" loses. Do the back-of-envelope: tokens × calls × price, p99 latency, where you'd cache or tier models.
- Reaching for LangChain by reflex. Being able to write the agent loop yourself (and say why you would or wouldn't pull in a framework) reads as more senior than naming tools.
Page checkpoint
Did the AI-eng mock prep stick?
Pass to unlock the Next button belowWhat's next
→ Pair this with the AI system-design interviews page, then run a round at SoloMock (AI Engineer track). For the general coding loop, the SWE Interview Guide shares the same problem set.