Simulation rules
- Use prompts the candidate has not rehearsed verbatim.
- The observer must interrupt, clarify, and push back as a real interviewer would.
- Score each round before discussing it; do not coach during the round.
- Take 5–10 minute breaks, but do not read notes between rounds.
- Turn misses into at most three repair tasks. A list of 20 problems is not a diagnosis.
Quick builder
Generate a concrete prompt set
Select a role to draw one core-library prompt for each typical round. Regenerate to change prompts; replace any question the candidate rehearsed recently.
Role-specific packet
Applied Scientist
3h 35m including breaksPrefer an experienced scientist, MLE, product partner, or interviewer.
- 01
ML breadth 45 min
Choose three unseen breadth questions. Require mechanism, failure mode, and one alternative on each.
- 02
ML system design 45 min
Design an ML system under explicit user, latency, cost, and delayed-label constraints.
- 03
Project deep-dive 40 min
Probe ownership, the hardest decision, one failure, evidence, and what the candidate would change.
- 04
Product & experimentation 35 min
Define a ship decision, primary metric, guardrails, experiment unit, and validity threats.
- 05
Behavioral / leadership 30 min
Ask one conflict and one prioritization question; push back on at least two claims.
Role-specific packet
Machine Learning Engineer
3h 40m including breaksPrefer a strong software engineer or MLE who will inspect correctness and operational detail.
- 01
ML implementation 45 min
Use an ML primitive, evaluation, or debugging problem. Require examples, tests, complexity, and working code.
- 02
ML system design 45 min
Design data, training, serving, monitoring, rollback, and iteration for a product ML system.
- 03
Systems / production 45 min
Diagnose a training or serving bottleneck from incomplete symptoms; require a measurement plan.
- 04
ML breadth 35 min
Use two mechanism questions plus one model-choice trade-off.
- 05
Behavioral / project 30 min
Probe a production failure, ownership boundaries, and influence during incident recovery.
Role-specific packet
Research Engineer
3h 50m including breaksPrefer an RE, systems engineer, or researcher who can test both implementation and evidence quality.
- 01
ML implementation 45 min
Implement an ML primitive or debugging task with tests and complexity analysis.
- 02
Training / inference systems 45 min
Quantify a workload, identify the bottleneck, select parallelism, and design recovery.
- 03
Research depth 45 min
Critique a claim, identify alternative explanations, and design discriminating ablations.
- 04
ML breadth 35 min
Test mechanisms that connect algorithms to training or systems behavior.
- 05
Project deep-dive 40 min
Probe research-to-production translation, a failed experiment, and the candidate’s personal decisions.
Unfamiliar transfer prompt pool
Use one prompt the candidate has not seen. The goal is transfer of framing, evaluation, and trade-off patterns, not recall of a memorized architecture.
- A company has historical project documents and wants to predict which future ML projects will succeed. Scope the problem before proposing a model.
- Your model improves the offline metric by 8% but hurts the online primary metric by 2%. Decide what to do next.
- Inference quality is fixed; cut annual serving cost by 60% without violating the p95 latency SLO.
- A new policy forbids storing raw labels after 24 hours. Redesign training, evaluation, and auditability.
- Two teams report opposite experiment results for the same model. Design the investigation.
- The strongest paper baseline is missing and the reported gain is within seed variance. Decide what evidence you need.
Round scorecard
Score 0–2 before giving feedback: 0 = missing or unsafe, 1 = workable but inconsistent, 2 = strong under follow-up.
Go / risk / extend decision
Green
No critical dimension scores 0; every round has a workable structure; follow-ups reveal real depth; remaining misses are bounded.
Amber
One critical round is inconsistent but responds to targeted repair. Keep the loop only if there is time for a spaced retry before it.
Red
More than one critical round scores 0, ML implementation or design cannot reach a baseline in time, or stories lose ownership under follow-up. Extend or move the loop when possible.
After the simulation
- Write the failure as an observable behavior, not “study systems more.”
- Choose one corrective drill and a due date.
- Retry closed-book after spacing.
- Graduate only after the repair survives a mixed session.