AI engineering interview practice, beyond the first diagram
The template gets you ten minutes. Practice the other forty.
The CTO wants a bigger model by Friday
Your RAG assistant answers questions for a law firm. Spot checks find it wrong on 8% of legal queries. The CTO wants to switch to a bigger model by Friday.
- 40k queries a week
- Bigger model costs 4x per query
- Today is Monday
Preview: the follow-up is chosen from lines written in advance, based on what you wrote. Nothing you type is stored.
Interview practice for AI engineers and forward deployed engineers. You give your answer, the interviewer pushes on it, and you revise until the reasoning holds up.
9 free chapters · 32 practice questions · 9 interview rounds mapped for AI Engineer and FDE
What the template misses
Most system design prep ends once the boxes are on the board. In a real interview that is roughly where the questions that decide the outcome begin.
The diagram every course teaches.
The interviewer circles one box.
You pick a path and defend what it costs.
Every course teaches frame one. We train frames two and three.
Know the loop before you walk in
The rounds AI Engineer and FDE candidates usually face, in order, with what each one tests and whether you can use an AI assistant in it.
01 · AI off
Coding fundamentals
Python, data structures, parsing and clean code.
45-60 min live
Not drilled here. Good coding practice already exists.
02 · AI on
Working with AI
Direct a coding assistant, review its diff, and extend a small repo with tests.
60-120 min
03 · AI off
LLM fundamentals
Tokens, context, sampling, embeddings, and when to prompt, retrieve or fine-tune.
45 min conceptual
04 · AI off
GenAI system design
RAG, agents, guardrails, cost and latency, and what breaks at 10x.
60 min whiteboard
05 · AI off
ML system design
Thresholds, metrics that move the business, and when a rule beats a model.
45-60 min whiteboard
06 · AI off
Evals & debugging
Design an eval, catch LLM-judge bias, and trace a wrong answer to its cause.
45-60 min
07 · AI varies
Take-home assignment
Build a scoped task, then defend what you chose, what you left out and what you would change.
A few hours to build, then a 45-60 min defence
Round shapes come from company guides and candidate reports, checked October 2026. Loops vary by company and team, so ask your recruiter which rounds you will get and whether AI tools are allowed in each.
How a scenario works
One rung from a real scenario, as you would play it. The interviewer reacts to what you wrote, then the rung is scored against a four-level rubric that quotes your own words back to you.
ML system design
Set a fraud threshold when a missed fraud costs 20x a false block
You answer
You run fraud ML at a payments company. Every card transaction gets approve, block or review inside the checkout call, and the model you inherited scores each one.
- p99 ≤ 200ms for the decision
- fraud ≈ 0.1% of transactions
- false block costs $50-200
- missed fraud costs $500-5,000
A legit customer gets blocked and a fraudster gets through. Which mistake hurts more, and what does that do to where you put your threshold?
A missed fraud costs far more than blocking a good customer, so I'd keep the threshold low. Anything borderline goes to manual review.
The interviewer pushes back
Put a number on it. Where does the line go, and why there?
You named the cost gap but never put a number on the line, so the next question asks for exactly that. A stronger answer would have moved straight on to the next rung.
Your rung is scored
Trade-off
Level 2 means: Names that a missed fraud costs more than a false block but doesn't turn it into a rule.
Evidence: “A missed fraud costs far more than blocking a good customer”
Naming the cost difference is the first half. The second is turning it into a rule: with average costs of $100 and $2,000, block when the chance of fraud times $2,000 is more than the chance it's legitimate times $100, which puts the line near a 5% fraud score.
Revise your answer. The revision is scored too.
Learn the reasoning first
Every lesson is free. Practice is where we charge.
MLOps Production Guide · 01
ML Problem Framing
Start with the business outcome, then work top-down: product outcome -> model goal -> decision policy. If those four layers do not line up, no model improvement will save the system.
30 min read · 4 exercises
MLOps Production Guide · 02
MLOps Blueprint & Operational Strategy
MLOps is the operating system for ML in production — code, data, and models all change independently, so you need explicit pipelines, observability, and retraining loops for each.
65 min read · 5 exercises
MLOps Production Guide · 03
MLOps Platforms
Platform choice is an architectural decision — match it to your team's maturity and workflow, not to the vendor's feature list.
40 min read · 3 exercises
Who builds this
Deepak Karkala
- Independent review
- Every chapter and scenario goes to a separate AI reviewer that checks the technical claims. Its fixes go in before anything ships.
- A human signs off
- Nothing goes live until a person has read the reviewed version.
- Sources and assumptions
- Numbers in a scenario are labelled as assumptions. A claim about a specific company carries its source and the date it was checked.
- Grader accuracy
- We are measuring the grader against human graders and will publish the agreement rate here once we have one.
Pricing
No subscription. Pick your role, prepare for one to three months, get the job.
Free
$0
- Every chapter, with its exercises and apply-its
- Two complete practice scenarios in every round
- One run of your role's first mock interview, at Mid-level
- Your progress across rounds and skills
Pro
30 days
$29 for 30 days
One payment. Never renews.
90 days
$59 for 90 days
One payment. Never renews.
- Everything in Free
- Every practice scenario in every round
- Both mock interviews for each role at Mid-level, Senior and Staff, as they launch
- Single-round mocks before a real interview
- Expert answers and AI feedback on every practice question
No subscription: one payment, never renews. Full refund within 7 days (30-day pass) or 14 days (90-day pass), even after use, once per person. Refund policy
Questions
Why not ask ChatGPT to interview me?
You can, and it helps for breadth. A chat model asks whatever comes to mind and tends to accept a confident answer. Our follow-ups are written and reviewed in advance, each one aimed at a specific gap, and your answer is scored against a rubric you can read.
How accurate is the scoring?
An AI grader scores each answer against the question's rubric and shows which parts you covered. It can be wrong, so Pro also shows the expert answer to compare against. We are measuring the grader against human graders and will publish the agreement rate once we have it.
Do companies actually ask this?
The round shapes come from company guides and candidate reports, and loops vary by company and team, so check with your recruiter. We only say a question was reported at a company when we can cite where.
AI-on or AI-off?
Both exist now. Design and conceptual rounds are usually AI-off. A growing number of coding and build rounds allow an assistant and grade how you direct and check it. The round map above marks which is usual for each round.
What's free and what's paid?
Free: all 9 chapters with their exercises, two complete practice scenarios in every round, one run of your role's first mock interview at Mid-level, and your progress. Pro ($29 for 30 days or $59 for 90 days, no subscription) adds every other scenario, mock interviews at every level, single-round mocks, and expert answers for all 32 questions in the practice bank.
Can my employer pay?
Yes. Pro is $29 for 30 days or $59 for 90 days, billed to any email address, so a corporate card works. Each pass is a single payment that never renews, which is easy to expense.
What if it isn't right for me?
Ask within 7 days if you bought 30 days, or 14 days if you bought 90 days, for a full refund, even if you used it (once per person). Nothing renews, so there is nothing to cancel.
And then what?
Get the next scenario by email when it ships: one decision, the follow-up that breaks the obvious answer, and what a strong revision looks like.