FDEBench Join us live

Prepare for the interview.

159 interview-style questions for forward deployed and AI engineering roles. 22 come with the full scoring rubric, and 23 are linked to a named company by a candidate report or prep guide.

What the loops ask

Showing all of 159 questions

Each question opens on its own page, with the full rubric or a reference sketch where one exists. 71 come from FDEBench’s own bank; the rest from candidate reports, prep guides and videos (IGotAnOffer, the AI Engineering Field Guide, Keerti Purswani, DesignGurus and others). FDEBench is not affiliated with any company named.

What the loops ask

From one student’s eight AI interview loops

What they ask most

Customers and stakeholders

How you handle people who aren’t technical, and whether you’ve worked with outside customers.

The technical rounds

They vary a lot

Debugging real code and logs, or building a feature from scratch, some with AI and some without.

The final round

A pitch, under questions

A deck and a proof of concept built on their product, presented like a customer or investor pitch.

If you come from engineering, expect them to test the customer side hardest.

A bootcamp student, October 2026, shared with permission and paraphrased.

One framework for each round

A framework is how a 45-minute room hears judgment

  • Customer and stakeholder questions

    STAR: situation, task, action, result

    “Here’s what was going on, what I had to do, what I did, and what changed.” End on a number.

  • Debugging and “fix this”

    Reproduce it, read the logs, change one thing

    “I can make it fail every time. The logs point here. I changed one thing. Next I’d check this.”

  • AI system design

    The seven moves, below

    “Before I draw anything: what number would tell us it works?”

  • The pitch and the proof of concept

    The five-line brief

    “What it does, what it costs per run, what could go wrong, who signs off, and the one number we’d track.”

The design round, in seven moves

43 of the 45 minutes, and what you say at each

Contract 5 minminutes 0–5

“Before I design anything: what does good look like? How fast, how accurate?”

Baseline 2 minminutes 5–7

“Simplest version: one prompt, one model call. Here’s where it falls short.”

Flow 12 minminutes 7–19

“Let me walk one request from the moment it comes in to the reply.”

Eval 8 minminutes 19–2716.5% of points

“Here’s how I’d know it works before a single user complains.”

Failure 8 minminutes 27–3531% of points

“Here’s what breaks first, and what the user sees when it does.”

Numbers 5 minminutes 35–40

“Let me do the cost out loud: tokens per call, times calls, times price.”

Iterate 3 minminutes 40–43

“In week one I’d ship this small piece, measure it, then add the rest.”

47.5% of the points sit in eval and failure.

In FDEBench’s first 10 rubrics (63 criteria worth 1,000 points), failure modes carry 31% of the points and evaluation 16.5%.

Show the seven moves as a table
MoveMinutesStarts atShare of points
Contract50–
Baseline25–
Flow127–
Eval81916.5%
Failure82731%
Numbers535–
Iterate340–

A worked answer

Illustrative numbers

The question“Your AI agent keeps calling the wrong tool. How do you fix it?”
  1. ContractThe right tool on the first try for 95 of 100 real requests.
  2. BaselineRead 50 traces, the log of each request, and count which tools get confused.
  3. FlowRequest → tool descriptions → the choice → its arguments.
  4. Eval100 logged requests labelled with the right tool, scored before and after every change.
  5. FailureTwo tools with overlapping descriptions, too many tools, a vague request: ask a question back.
  6. NumbersFewer tools means a shorter prompt and a lower bill.
  7. IterateRewrite the descriptions first, then cut or group tools, then add routing: a first step that picks the group.

Start with the cheapest fix you can measure.

Guides for each round

All guides

Questions people ask

What do forward deployed engineer interviews ask?

Across one student’s eight AI interview loops: customer and stakeholder questions came up most, the technical rounds varied (debugging real code and logs, or building a feature, with or without AI), and the final round was a pitch: a deck and a proof of concept built on the company’s product. What the loops ask.

How do I prepare for an FDE interview?

Use one framework per round. STAR for customer and stakeholder questions, “reproduce it, read the logs, change one thing” for debugging, the seven moves for AI system design, and a five-line brief for the pitch. Then practice out loud on real questions and score yourself against a rubric. The interview, round by round.

What is in a full rubric?

22 questions have one. It lists what a strong candidate asks before designing, criteria weighted to 100 points with what earns full credit and the feedback if you miss it, and the red flags that would end the interview.

Where do the questions come from?

71 come from FDEBench’s own bank. The rest come from candidate reports, Reddit, prep guides, field guides, newsletters and YouTube. 23 are linked to a named company, and each says whether a candidate reported it or a guide names it.

How should I spend a 45-minute AI system design round?

In seven moves: agree what good looks like (5 minutes), the simplest baseline (2), the request flow (12), how you would evaluate it (8), what breaks first (8), the cost out loud (5), and what you would ship in week one (3). In FDEBench’s first 10 rubrics, failure modes and evaluation carry 47.5% of the points.

Want feedback from a live cohort?

The bootcamp includes a module on landing the role and passing the interview. The next cohort starts 14 Nov.

Join us live