Skip to content

Interview Field Reports

nanorag edited this page Sep 1, 2026 · 1 revision

Interview field reports

What people were actually asked, and how it went. Anonymised, community-contributed, and more useful than any list of predicted questions — because the gap between what a company says it tests and what it tests is where preparation goes wrong.

This page lives in the wiki rather than in docs/ deliberately: it should be editable by whoever just came out of a loop, without a pull request and without review.


What to include

Do include: the round shapes, the kind of questions, what surprised you, what you wish you had prepared, and the follow-up that caught you out.

Do not include: the company's name if you signed anything saying not to, verbatim take-home prompts, or anything you were told was confidential. A round shape is not a leak; a verbatim problem statement usually is.

Use the shape below. Keep it honest about what went badly — a report where everything went well teaches nobody.

### YYYY-MM · Role shape · outcome

**Rounds:** screen → technical depth → system design → behavioural

**What they actually probed:**
- ...

**The follow-up I did not have ready:**
> "..."

**What I would prepare differently:**
- ...

Reports

2026-08 · Forward-deployed / deployment engineer · offer

Rounds: screen → case study (take-home, 4h) → system design → behavioural

What they actually probed: almost nothing about model internals. The case study handed over a messy corpus and a vague goal, and the scoring was whether you built an eval set before you built anything else. Two of the three people I compared notes with went straight to improving retrieval and did not pass.

The follow-up I did not have ready:

"How many questions would your eval set need to detect the improvement you are claiming?"

I said "a few hundred" and could not defend it. The power arithmetic takes ten minutes to learn.

What I would prepare differently: rehearse saying "week one is the eval set" out loud. It feels like stalling when you say it and it is the highest-scoring sentence in that round.


2026-07 · ML engineer, search team · rejected at system design

Rounds: screen → coding → system design → hiring manager

What they actually probed: the coding round was ordinary. The design round was entirely about what breaks at scale, and I had prepared architecture rather than failure modes.

The follow-up I did not have ready:

"You said you would filter by permission. Where does that filter run?"

I said "after retrieval" and watched the interviewer's face change. Post-filtering collapses k unpredictably and leaks — the result count is a function of content the user cannot see. It is in The Inference Channel Audit and I had not read it.

What I would prepare differently: for every component, know the failure mode, not just what it does. They ask about failure.


2026-06 · Applied research · offer, declined

Rounds: paper discussion → experiment design → coding → team fit

What they actually probed: the paper round was not "do you know this paper". It was "design the experiment that would show they are wrong." Falsification, not comprehension.

The follow-up I did not have ready:

"What result would make you believe them?"

A critique you cannot satisfy is not a critique, and I had only prepared attacks.

What I would prepare differently: for any technique, be able to state its precondition — what failure it fixes — and how you would check whether you have that failure. That single habit answered about half of this loop. It is The Precondition Test.


Patterns across reports so far

Three, from a small sample — treat as hypotheses, not findings:

  1. The eval set is the differentiator in FDE-shaped loops. Candidates who tune before measuring do not pass, and they usually do not know why.
  2. Design rounds ask about failure, not architecture. Preparing a clean diagram prepares you for the wrong twenty minutes.
  3. The second follow-up is where it is decided. Everyone survives the first. Level 3 and 4 of the question decoder is where offers separate.

Add yours. Three reports is an anecdote; thirty is a signal.

Clone this wiki locally