Is coding still tested at ML interviews?

Yes, almost always. The bar is slightly lower than a pure SWE loop but high enough to filter — a bad coding round closes the loop regardless of ML strength.

What's the most under-prepped round?

The ML system design round. Candidates rehearse model choice and skip the framing (labels, features, feedback loops, monitoring). Framing is where Staff-level signal lives.

Do I need to know the latest models (GPT-5, Claude Opus 4.7, etc.) cold?

Know the landscape, but don't over-index. Interviewers reward clear tradeoff reasoning over latest-model name-dropping.

How do applied scientist and ML engineer interviews differ?

Applied scientist loops lean on research depth and statistical rigor. ML engineer loops lean on production concerns (serving, monitoring, retraining). Research depth is a dial that shifts between them.

Should I prep for the research round even if I don't have a PhD?

Yes — but the 'research round' for a non-PhD is often a deep-dive on a production ML project you've built. Have 2–3 projects ready with measurable outcomes, model iterations you tried, and what didn't work.

Guide · ML

ML engineering interview — coding, ML system design, modeling, and research.

TL;DR

ML engineering loops layer four distinct rounds: standard coding, ML system design, modeling/hands-on ML, and a research-depth or production-ML round. The coding round is nearly identical to SWE loops. The ML rounds reward clear framing of labels, features, and feedback loops over cutting-edge model choices.

Who this is for

Candidates targeting ML engineer, applied scientist, or ML platform roles.

Time commitment

12–16 weeksof structured prep. Less if you've been interviewing recently; more from a cold start.

Round 1: Standard coding

Treat this as an SWE round. Same 12 patterns, same rubric. A common trap: ML candidates under-prep this round because they assume the ML rounds carry more weight. They don't; a fumbled coding round filters you out before the ML rounds happen.

Round 2: ML system design

Design a recommendation system, a ranking service, a fraud detector. The rubric is data → features → labels → model → serving → feedback loop → monitoring. The model choice is the least interesting part — framing is load-bearing.

Round 3: Modeling / hands-on ML

A dataset, a notebook, and 60–90 minutes. Ask what the label is, what the eval metric is, what the baseline is, and why. Most candidates jump to XGBoost; strong candidates articulate the baseline first.

Round 4: Research depth or production ML

Depending on the team: either a deep-dive on a paper or past project, or a round on serving, monitoring, and retraining. For applied scientist roles it's the former; for ML platform roles, the latter.

Patterns to prioritize

Drill these first. Each links to a dedicated pattern page with template, scenarios, and reference code.

Dynamic ProgrammingBreak an overlapping-subproblem problem into a recurrence and cache results.Graph BFS / DFSBreadth-first for shortest unweighted paths; depth-first for exhaustive traversal.Heap & Priority QueueA data structure that returns the min or max in O(log n) per operation.Prefix & SuffixPrecompute cumulative arrays so range queries answer in O(1).Two PointersTwo cursors moving independently over a sorted or monotone structure.

Frequently asked questions

Is coding still tested at ML interviews?: Yes, almost always. The bar is slightly lower than a pure SWE loop but high enough to filter — a bad coding round closes the loop regardless of ML strength.
What's the most under-prepped round?: The ML system design round. Candidates rehearse model choice and skip the framing (labels, features, feedback loops, monitoring). Framing is where Staff-level signal lives.
Do I need to know the latest models (GPT-5, Claude Opus 4.7, etc.) cold?: Know the landscape, but don't over-index. Interviewers reward clear tradeoff reasoning over latest-model name-dropping.
How do applied scientist and ML engineer interviews differ?: Applied scientist loops lean on research depth and statistical rigor. ML engineer loops lean on production concerns (serving, monitoring, retraining). Research depth is a dial that shifts between them.
Should I prep for the research round even if I don't have a PhD?: Yes — but the 'research round' for a non-PhD is often a deep-dive on a production ML project you've built. Have 2–3 projects ready with measurable outcomes, model iterations you tried, and what didn't work.

Run the free diagnostic.

Ten-minute patterns quiz. No card. Personalized loop starts on the other side.

Start the diagnostic