Company interview guide
Amazon Machine Learning Interview Guide (2026)
Leadership Principle questions in every round, standard algorithm coding, and ML breadth, depth and design rounds that differ by track.
Amazon's machine learning interview loop has 8 reported stages, from recruiter call to bar raiser & leadership principles; typical length: 2-6 weeks.
Roles covered: ML Engineer, Applied Scientist and Research Scientist. The question bank holds 58 questions reported from Amazon interviews, 11 of them free to solve in your browser.
- Interview process length
- 2-6 weeks
- Stages
- 8
- Reports linked to bank questions
- 59
- Bank report dates
- Mar 2020–Sep 2026
- Questions in bank
- 58 (11 free)
- Last reviewed
- October 2026
How the Amazon interview loop works
Most reports run: a recruiter call, an online assessment for many ML Engineer and intern applicants, one or occasionally two technical screens of 45-60 minutes, then a loop of four to six interviews of about an hour, held on one day or split over two. Amazon's own science prep page describes one or two 60-minute screens and a loop of four 55-minute interviews, with an outcome within five business days. Some loops include a hiring manager conversation, and science candidates may be asked for a tech talk. Several candidates say the prep email did not match the real rounds, so prepare for the full menu.
The Amazon interview process, stage by stage
01Recruiter call
Usually about 30 min by phone or video
A short conversation about your background, the team and the level. Recruiters often follow up with a prep email listing the round types; treat it as a guide rather than a promise, since several candidates found the real rounds differed.
What it tests
- How your experience maps to the specific team's job description
- Motivation for Amazon and for the role
- Logistics: level, location, timeline
How to prepare
- Ask which loop competencies apply (coding, ML breadth, depth, application or design, tech talk) and whether there is an online assessment.
- For research roles, ask whether a PhD is expected; one candidate passed round one and was then stopped for lacking one.
- Have a two-minute summary of your most relevant project and a first Leadership Principle story ready.
Recruiter or HR call before the technical rounds in 22 ML-role candidate reports (16 since 2024), usually about 30 minutes; the company's science prep page only mentions a recruiter arranging the loop.
02Online assessment
Timed online test: usually 2 coding problems in about 90 min; some versions add a work simulation and work-style survey, intern versions add ML multiple choice
Mostly an ML Engineer and intern step: 15 of the 19 reports that mention it come from engineering or intern applicants, while Applied and Research Scientist candidates usually go straight to a technical screen. Problems are standard data-structure questions, and several candidates found the second problem the real filter.
What it tests
- Solving two medium (occasionally hard) algorithm problems under time pressure
- Passing hidden test cases, including output ordering and edge cases
- Judgement in situational work-simulation questions
- ML concept multiple choice on some intern and program assessments
How to prepare
- Practise timed sets of two problems in 90 minutes; candidates often call the second problem much harder.
- Read output format rules carefully, since one candidate failed a test case only because of result order.
- Answer the work-style survey consistently with the Leadership Principles.
Online coding test in 19 ML-role reports, 15 of them ML Engineer or intern; only 7 since 2024 and 4 from science candidates; the company's pages describe a coding plus work-style OA for software and university science roles.
03Technical screen
1 (sometimes 2) x 45-60 min video call with a scientist, engineer or hiring manager on the team, shared editor
A compressed version of the loop: a Leadership Principle question or two, a walk through your past work, quick ML questions and, in most reports, one easy to medium coding problem. One reported split was 15 minutes behavioral, 15 minutes ML fundamentals and 25 minutes of coding. Some science screens are ML and project discussion only, and one ML Engineer screen with a hiring manager was a low-level design of a sales-rank leaderboard.
What it tests
- Explaining one project end to end and defending each design choice
- ML basics on demand: metrics, losses, normalization, attention, perplexity
- One easy or medium coding problem written cleanly while talking
- One or two Leadership Principle stories with measurable results
How to prepare
- Keep ML answers short and precise; the interviewer is trying to cover many topics in a few minutes.
- Expect a grid, tree, binary search or dynamic-programming problem; explain the approach before you type.
- For recommender or ranking teams, be ready to reason about CTR models, sparse features and what to do after an experiment shows a lift.
Free reported questions for this stage
- Transformer Attention and the 1/sqrt(d_k) ScalingLLMs & GenAIMediumSolve
- Reading a 2% CTR Lift in a Music Recommendation TestA/B Testing & CausalHardSolve
Live technical screen in 44 ML-role reports (30 since 2024); most mix project talk, ML questions and one coding problem, matching the company's description of 60-minute screens.
04Loop: coding
ML Engineer: usually 2 x 60 min of algorithm coding; science: 1-2 x 60 min, sometimes ML coding; shared editor, code often not run
Mostly classic easy to medium algorithm problems, with trees, graphs, grids, monotonic stacks and dynamic programming the most reported themes; one engineering candidate notes that ML Engineers are held to the software engineer coding bar. Science loops and LLM teams sometimes swap a problem for ML code written from scratch, such as sparse cosine similarity with top-k, Chamfer distance, or attention with a KV cache and GRPO in numpy.
What it tests
- Tree and graph traversal, reconstruction and search
- Choosing data structures that give logarithmic updates and queries
- Writing syntactically complete code without an IDE or a run button
- Implementing similarity and distance computations without library shortcuts
- For LLM teams: multi-head attention, KV caching and RL fine-tuning maths in numpy
How to prepare
- Practise medium tree and graph problems until you can code them in 25 minutes, leaving room for the behavioral part of the hour.
- State complexity and test a small example by hand, since the code may never be executed.
- If the team is ML-heavy, rehearse writing cosine similarity, kNN and attention from scratch in plain Python and numpy.
Free reported questions for this stage
- Pairwise Cosine Similarity in NumPyML CodingEasySolve
- Multi-Head Causal Self-Attention From ScratchML CodingHardSolve
Coding rounds in 25 ML-role loop reports (19 since 2024): ML Engineer loops usually two algorithm rounds, science loops one or two, with ML-flavoured coding in about 6; the company's prep pages list coding in the loop.
05Loop: ML breadth
About 60 min of rapid conceptual questions plus Leadership Principle questions; a named round mainly in science loops
A wide quiz across classical ML, deep learning, statistics and, increasingly, LLMs and agents; one recent candidate called it very extensive. Questions often probe what changes when a standard choice is swapped, such as training a classifier with MSE or tying the attention weights. Science loops almost always include this round, while recent ML Engineer loops more often fold ML questions into the design rounds.
What it tests
- Metrics for imbalanced classification and when to prefer each
- Loss functions and their gradients: cross-entropy, MSE, KL
- Regularization, bias-variance, cross-validation and tree ensembles
- Optimization: gradient descent versus closed form, SGD, Adam, convexity
- Modern topics: transformer internals, choosing an LLM, defining tools for an agent
How to prepare
- Answer each question in two or three sentences, then offer to go deeper; breadth rewards coverage.
- Be ready for what-if variants (swap the loss, tie the attention weights, remove the hidden layers).
- Research and economics roles add probability and experiment design, such as confounding or the newsvendor problem.
Free reported questions for this stage
- MSE or Cross Entropy: Picking a Loss for a Probabilistic ClassifierClassical MLMediumSolve
- Why L1 Regularization Produces Sparse SolutionsClassical MLEasySolve
- Logistic Regression: Gradient, Hessian, and ConvexityStatistics & InferenceMediumSolve
Named ML breadth round in 16 loop reports: 11 science, mostly 2024-2026, and 5 ML Engineer from 2020-2024; breadth questions also fill many screens; science prep pages list depth and breadth as loop topics.
06Loop: ML depth & tech talk
About 60 min on your own work and specialty; some science loops add a 60-min tech talk on a topic you choose
The interviewer picks a project or paper from your resume and goes as deep as it will go, then asks about the underlying techniques in your field. Reported examples range from vision transformers and triplet loss for an image-similarity project to research directions beyond CLIP, and distributed training internals for infra teams. Candidates warn that anything you mention becomes fair game.
What it tests
- Why you chose each model, loss and data decision, and what the alternatives were
- First-principles knowledge of the methods behind your project
- Systems depth for infra teams: distributed training primitives and parallelism
- LLM inference cost and memory trade-offs
- Presenting and defending your research in a tech talk
How to prepare
- Pick two projects and prepare every follow-up: data, baselines, metrics, failures and what you would change.
- Do not list a technique on your resume unless you can derive or implement it.
- For a tech talk, pitch it to scientists outside your subfield and leave time for questions.
Free reported questions for this stage
- What the KV Cache Stores and Why It Speeds Up DecodingLLMs & GenAIMediumSolve
- What torch.distributed.barrier Does and How It FailsDeep LearningMediumSolve
Project or research deep-dive round in 16 loop reports, mostly science track; a tech talk appears in 4 reports and is listed on the company's science prep page.
07Loop: ML application & system design
Science: usually 1 x 60 min science application on a team problem; ML Engineer: 1-2 x 60 min ML system design
An open-ended problem to solve end to end: data, labels, features, model, metrics, deployment and risks, usually close to the team's own product. Reported prompts include dark-store demand forecasting, metrics to filter crowdsourced training data, search results, a video recommender, text-to-image search and anomaly detection over logs. ML Engineer loops on LLM and infra teams have asked for a managed LLM fine-tuning platform on AWS and an LLM chatbot with reward design.
What it tests
- Turning a vague business goal into an ML problem with a clear objective
- Defining offline metrics and an online experiment
- Data collection and label quality at scale
- Serving, cost and multi-tenant concerns for LLM platforms
- Customer impact and trade-offs, framed in Leadership Principle terms
How to prepare
- Read the job description and the team's public work; the prompt is often their real problem.
- Start with the customer and the success metric before choosing a model.
- Have one ready blueprint each for forecasting, search or recommendation, and an LLM platform.
Free reported questions for this stage
- Designing a General-Purpose LLM Chat AssistantLLMs & GenAIHardSolve
Design or science-application round in 21 ML-role loop reports (18 since 2024), often LLM or infra design; the company's pages describe system design for software roles and experimental design for research science, not a named round.
08Bar Raiser & Leadership Principles
Two or three Leadership Principle questions in every round, plus a Bar Raiser interview from outside the team that is often fully behavioral
Amazon scores its 16 Leadership Principles throughout the loop, and two candidates say behavioral questions took half or more of each technical hour. The Bar Raiser, an interviewer from outside the hiring team, mostly asks behavioral questions and helps run the debrief. Expect follow-ups on exactly what you did, how you measured it and what you would do differently; one candidate on an AGI team saw fewer behavioral questions than usual.
What it tests
- Specific past examples for principles such as Customer Obsession, Ownership, Dive Deep and Deliver Results
- Disagreement and commitment, failures and lessons learned
- Delivering under tight deadlines and deciding with incomplete data
- Quantified impact and your personal contribution
How to prepare
- Prepare at least one STAR story per principle and two for the most common ones; reuse is noticed.
- Put numbers in every result and say I, not we, when describing your part.
- Include a genuine failure story and one about feedback you received; both come up often.
Leadership Principle questions in 50 ML-role reports, two saying they took half or more of each round; a Bar Raiser round is named in 4 and described on the company's hiring pages.
Questions and themes reported most often
- Leadership Principle stories in every round: tell me about a time when (50 reports, bar raiser & leadership principles)
- Deep dive on a past project or paper, with every decision questioned (30 reports, loop: ml depth & tech talk)
- Tree, graph or grid coding: islands, rotting oranges, BST from preorder, next greater element, LCA (14 reports, loop: coding)
- Open-ended science application or ML design based on the team's product (12 reports, loop: ml application & system design)
- Transformer internals: attention, KV cache, RoPE, long-context models (9 reports, loop: ml breadth)
- LLM and agent questions: prompt evaluation, RAG, choosing an LLM, agent tools (5 reports, loop: ml breadth)
- Tech talk presenting your own research (4 reports, loop: ml depth & tech talk)
- Loss functions: cross-entropy versus MSE, cross-entropy formula, KL divergence (4 reports, loop: ml breadth)
What Amazon asks about
- Behavioral & Leadership Principles25%
- General coding23%
- ML breadth & concepts18%
- ML application & system design17%
- Project deep dive & ML depth11%
- ML coding6%
Amazon interview prep checklist
- Have at least one quantified STAR story for each of the 16 Leadership Principles, including a failure
- Solved a set of tree, graph and DP medium problems in under 30 minutes each, without running the code
- Can answer 30 rapid ML breadth questions in two or three sentences each
- Prepared a deep dive on two projects, with alternatives, metrics and lessons
- Can sketch an end-to-end design for a forecasting, search or recommendation problem with offline and online metrics
- Wrote cosine similarity, kNN and attention with a KV cache from scratch
- Know which org the team sits in, which track the role is on and what its likely application problem is
Tips
- Treat Leadership Principles as a full subject: write STAR stories with numbers, map them to the 16 principles and rehearse follow-ups.
- Budget up to half of each technical hour for behavioral questions, and practise solving a medium coding problem in 25 to 30 minutes.
- Drill trees, graphs, grids, monotonic stacks and simple dynamic programming; these dominate reported coding questions on both tracks.
- On the science track, build a one-page ML breadth sheet covering metrics, losses, regularization, ensembles, optimizers and transformer basics, and practise short answers.
- On the ML Engineer track, expect two algorithm rounds and one or two ML system design rounds, and prepare for an online assessment.
- Prepare two projects for a deep dive, including the alternatives you rejected and what failed.
- Research the team: application prompts usually mirror its real product, and LLM, infra and forecasting teams test very different depth.
- Write code that would compile and run; several rounds use a plain editor where nothing is executed.
- Do not rely on the prep email alone; candidates report rounds that differed from it.
See the rest of the Amazon loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Amazon loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Amazon interview FAQ
How many interview rounds does Amazon have for ML roles?
Candidates report 8 stages: Recruiter call, Online assessment, Technical screen, Loop: coding, Loop: ML breadth, Loop: ML depth & tech talk, Loop: ML application & system design and Bar Raiser & Leadership Principles. Most reports run: a recruiter call, an online assessment for many ML Engineer and intern applicants, one or occasionally two technical screens of 45-60 minutes, then a loop of four to six interviews of about an hour, held on one day or split over two. Amazon's own science prep page describes one or two 60-minute screens and a loop of four 55-minute interviews, with an outcome within five business days. Some loops include a hiring manager conversation, and science candidates may be asked for a tech talk. Several candidates say the prep email did not match the real rounds, so prepare for the full menu.
How long does the Amazon ML interview process take?
Reported timelines run 2-6 weeks (2-6 weeks, sometimes longer).
What does Amazon ask in ML interviews?
By reported question mix: Behavioral & Leadership Principles 25%; General coding 23%; ML breadth & concepts 18%; ML application & system design 17%; Project deep dive & ML depth 11%; ML coding 6%. The most repeated themes: Leadership Principle stories in every round: tell me about a time when; Deep dive on a past project or paper, with every decision questioned; Tree, graph or grid coding: islands, rotting oranges, BST from preorder, next greater element, LCA.
How should I prepare for the Amazon ML interview?
Treat Leadership Principles as a full subject: write STAR stories with numbers, map them to the 16 principles and rehearse follow-ups. Budget up to half of each technical hour for behavioral questions, and practise solving a medium coding problem in 25 to 30 minutes. Drill trees, graphs, grids, monotonic stacks and simple dynamic programming; these dominate reported coding questions on both tracks.
Are there free Amazon interview questions?
Yes. LeetML has 58 questions reported from Amazon interviews, and 11 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- GoogleClassic algorithm rounds plus ML design, ML domain depth and Googleyness, judged by a hiring committee.
- MetaFast classic coding, an AI-assisted coding project and product ML design for recommendations and integrity.
- MicrosoftProject deep dives in most rounds, quick ML and LLM fundamentals, standard coding, from-scratch ML code and GenAI design, in a loop shaped by the hiring team.
- AppleTeam-run loops built on project deep dives, everyday algorithms, short from-scratch ML coding and product-flavored ML design.
- NVIDIATeam-run loops that mix algorithm coding, deep learning built from scratch and GPU-aware LLM performance.
- WalmartWide ML and statistics questioning, easy-to-medium coding and retail-flavoured ML design, run team by team.
Official sources
- Amazon: Applied Scientist Interview Prep (official)
- Amazon: Research Scientist Interview Prep (official)
- Amazon: University Applied Science Online Assessment Prep (official)
- Amazon: SDE II Interview Prep (official)
- Amazon: SDE II Online Assessment Prep (official)
- Amazon: Software Development Interview Topics (official)