Company interview guide
Zoox Machine Learning Interview Guide (2026)
Coding with language and OOP questions, rapid-fire ML and vision fundamentals, and an unusual math and reasoning round.
Zoox's machine learning interview loop has 8 reported stages, from recruiter call to behavioral, hiring manager and team chats; typical length: Not well reported; one candidate calls the process very long and a take-home review took about 10 days.
Roles covered: ML Engineer, Deep Learning Engineer and Data Scientist. The question bank holds 2 questions reported from Zoox interviews, 1 of them free to solve in your browser.
Owned by Amazon. Amazon acquired Zoox in 2020. Zoox runs as an independent Amazon subsidiary.
- Interview process length
- Not well reported; one candidate calls the process very long and a take-home review took about 10 days
- Stages
- 8
- Reports linked to bank questions
- 2
- Bank report dates
- May 2026–Jul 2026
- Questions in bank
- 2 (1 free)
- Last reviewed
- October 2026
How the Zoox interview loop works
Most reports start with a 30 to 45 minute recruiter call. Engineering candidates then have a technical screen of about an hour with an engineer, usually a coding problem plus questions on C++ or object-oriented programming; data science and some vision candidates were more often given a take-home first, though most of those reports date from 2022. The final stage is a virtual or in-person panel of about six interviews (one 2026 data science candidate counted nine): general coding, ML or computer vision fundamentals, a math and logical reasoning round, sometimes a technical presentation, and hiring-manager or cross-team conversations. One candidate was told interviews are weighed together but was rejected after one weak round. Several candidates say they never heard back after the panel, so follow up and keep other processes moving.
The Zoox interview process, stage by stage
01Recruiter call
30-45 min with an in-house recruiter
A conversation about your background, skills, past projects and logistics such as visa status. Some candidates were approached by a recruiter for a specific opening; for data science and vision roles the recruiter sometimes sends a take-home right after this call.
What it tests
- How your experience maps to the team's autonomy work
- Relevant projects in your current role
- Logistics: location, visa status, timeline
How to prepare
- Ask which rounds follow, how long each is and whether you will code in C++ or Python.
- Ask whether your track has a take-home, and when to expect it, so you are not left waiting.
Recruiter or HR intro call as the first step in 7 ML, data science and vision reports (four since 2024), described as 30 to 45 minutes on background, projects and logistics.
02Take-home assignment (data science, some vision roles)
A few hours to a few days, set by the recruiter; one deep learning candidate had a timed coding challenge
Data science reports most often begin with a take-home, such as a project on time-series and location data that took one candidate three to four hours. One vision candidate got several days to build an object detection approach, and one data science panel included a presentation of the take-home. Nearly all of these reports are from 2022, so ask whether your team still uses one.
What it tests
- Cleaning and modelling a real-looking dataset end to end
- Working with time-series and geographic data
- Building a working computer vision pipeline under a deadline
- Writing it up clearly enough to present later
How to prepare
- Make the code run out of the box and explain how to run it; one candidate suspected reviewers never executed theirs.
- Keep notes on your choices, since some panels ask you to present the take-home.
A take-home in 6 reports: four data science (three from 2022, one 2025 promised but never sent), one 2022 vision project and one 2022 timed deep learning challenge. No 2024+ ML engineer report has one.
03Technical screen
About 60 min on video or a coding platform with an engineer; some screens split into two 30 min problems
Usually one coding problem preceded by quick-fire questions about the language you use. Reports include C++ memory questions with a memory-efficient coding task, an object-oriented design exercise, two search and recursion problems in 30 minutes each, and in 2026 a parsing task that scores text records by type.
What it tests
- Clean, object-oriented code with sensible classes
- C++ details such as smart pointers and object lifetime
- Standard algorithm problems: search and recursion
- Parsing strings and applying per-type rules correctly
- Keeping memory use low and the solution scalable
How to prepare
- Revise the language you will use in depth (ownership and smart pointers in C++, classes and inheritance in Python).
- Expect short definition questions before the coding starts and answer them crisply.
- For rule-heavy parsing tasks, map each type to its rules in one place and check the output format exactly.
A technical screen in 4 ML engineer and vision reports from 2023 to 2026: language or OOP questions plus coding in two, a parsing task in one, two timed search problems in one. One data science report describes a similar mix.
04Panel: general coding
One or more 45-60 min rounds in the final panel; pseudocode was accepted in one 2025 loop
Standard algorithm and data structure questions, sometimes alongside a Python-specific round. One 2025 candidate had three one-hour parts (Python, an algorithm problem and ML basics) and was told code did not need to run.
What it tests
- Core data structures and algorithms
- Python fluency and idioms
- Explaining complexity and trade-offs
- Structuring a solution clearly even in pseudocode
How to prepare
- Practise medium-level algorithm problems in a plain editor, speaking your reasoning aloud.
- If running code is not required, spend the time on correctness and edge cases instead.
Coding in the final rounds in 5 reports: four ML engineer (2023 to 2025) and one 2022 data science panel; details are sparse beyond standard algorithm questions.
05Panel: ML and computer vision fundamentals
1-3 rounds of about 60 min, often question-and-answer; one 2025 senior loop had an ML design round
Interviewers check definitions, formulas and quick estimates, for example the self-attention formula and why both the softmax and the square-root scaling are needed, or the parameter count of a convolutional layer with 3 x 3 kernels. One 2023 candidate had three CV and ML rounds with overlapping questions; a 2025 senior candidate also had an ML design round.
What it tests
- Transformer attention: the formula, shapes and the role of the scaling
- CNN arithmetic: parameters, output sizes and compute
- Computer vision basics for perception tasks
- Core ML theory stated precisely
- Designing an ML system for the team's domain (senior roles)
How to prepare
- Memorize key formulas and be able to justify each term; candidates say exact definitions were expected.
- Drill quick parameter and shape calculations for convolution and attention layers.
- Answer the core point first and then add depth, since interviewers move on quickly.
Free reported questions for this stage
- Transformer Attention and the 1/sqrt(d_k) ScalingLLMs & GenAIMediumSolve
ML or CV rounds in 5 ML engineer reports (four since 2024): fundamentals and domain questions in four, ML design in one senior 2025 loop. Two say exact formulas were expected.
06Math and logical reasoning
One round of about 45-60 min in the panel; data science panels add statistics and probability
A round several candidates call unusual: short math, physics and logic questions, such as explaining why summer happens. One candidate felt the interviewer wanted specific keywords within a few minutes before moving on. Data science panels in 2022 had a math and problem-solving round and a separate statistics and probability round.
What it tests
- Explaining everyday physical phenomena from first principles
- Quick mental math and estimation
- Logic puzzles under time pressure
- Probability and statistics (data science)
How to prepare
- State the key mechanism in your first sentence (for seasons: the tilt of the Earth's axis), then elaborate.
- Refresh high-school physics and geometry alongside probability.
Math and reasoning round in 2 ML engineer reports (2023 and 2025) and math or statistics rounds in 2 data science reports from 2022.
07Technical presentation
Part of the panel in some loops: a talk on past work, or on the take-home for data science
One 2023 senior ML candidate gave a technical presentation at the start of the virtual onsite, and one 2022 data science panel included a presentation of the take-home. Ask your recruiter whether your loop has one.
What it tests
- Explaining a past project's goals, choices and results
- Defending methodology under questioning
- Presenting take-home findings clearly
How to prepare
- Prepare one project in depth, including what failed and what you would change.
- Keep the talk short enough to leave room for questions.
Only 2 reports mention a presentation: one 2023 senior ML engineer loop and one 2022 data science panel built on the take-home.
08Behavioral, hiring manager and team chats
1-2 conversations of 30-60 min, often with the hiring manager or a director; some candidates talk to other teams
Behavioral questions about past challenges plus conversations with the hiring manager and sometimes a manager from a partner team. One 2026 data science candidate ended with a director; one 2022 candidate was asked to speak with three more teams after references. In one 2023 loop the planned behavioral round became another ML round.
What it tests
- The most challenging thing you have worked through
- Collaboration with cross-functional teams
- Motivation for autonomous driving and for Zoox
How to prepare
- Prepare two or three stories about hard technical problems and your own role in them.
- Bring questions for each interviewer; candidates say time for them was often short.
Behavioral, manager or team-match conversations in 4 reports: one 2025 ML engineer, one 2026 and two 2022 data science; few give detail.
Questions and themes reported most often
- Self-attention formula: why divide by the square root of d, and why both softmax and scaling (1 reports, panel: ml and computer vision fundamentals)
- Count the parameters of a CNN layer with 3 x 3 kernels (1 reports, panel: ml and computer vision fundamentals)
- Parse text records and compute a score with per-type rules (1 reports, technical screen)
- C++ memory questions: keep a linked list of shared pointers alive (1 reports, technical screen)
- Object-oriented design exercise (airline ticketing) after language questions (1 reports, technical screen)
- Explain inheritance and polymorphism, then an OOP coding problem (1 reports, technical screen)
- Explain why summer happens (1 reports, math and logical reasoning)
- Classic expression calculator in a timed coding challenge (1 reports, take-home assignment (data science, some vision roles))
What Zoox asks about
- General coding & language questions35%
- ML & computer vision fundamentals25%
- Math & logical reasoning15%
- Take-home project (DS / CV)10%
- Behavioral & presentation10%
- ML design5%
Zoox interview prep checklist
- Can write the attention formula and explain the square-root scaling and the softmax
- Can compute parameters and output size of a convolution layer in under a minute
- Solved a rule-based string parsing problem with exact output formatting
- Revised C++ smart pointers and object lifetime, or Python OOP, for the screen
- Practised short first-principles answers to physics and logic questions
- Prepared a project deep dive that could become a technical presentation
- Asked the recruiter for the list, length and language of every round
Tips
- Expect knowledge checks: memorize core ML and vision formulas and be ready to justify every term.
- Ask early whether you will code in C++ or Python, and revise that language's details, not just algorithms.
- Practise quick parameter, shape and compute estimates for convolution and attention layers.
- Prepare for the math and reasoning round with short first-principles explanations of physics and geometry questions.
- Give the key point within the first minute of an answer; interviewers often move on fast.
- If you get a take-home, make it run with one command and document your decisions.
- Keep energy for a panel of about six rounds, and remember one weak round can decide the outcome.
- Follow up after each stage and keep other processes moving; several candidates heard nothing after the panel.
See the rest of the Zoox loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Zoox loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Zoox interview FAQ
How many interview rounds does Zoox have for ML roles?
Candidates report 8 stages: Recruiter call, Take-home assignment (data science, some vision roles), Technical screen, Panel: general coding, Panel: ML and computer vision fundamentals, Math and logical reasoning, Technical presentation and Behavioral, hiring manager and team chats. Most reports start with a 30 to 45 minute recruiter call. Engineering candidates then have a technical screen of about an hour with an engineer, usually a coding problem plus questions on C++ or object-oriented programming; data science and some vision candidates were more often given a take-home first, though most of those reports date from 2022. The final stage is a virtual or in-person panel of about six interviews (one 2026 data science candidate counted nine): general coding, ML or computer vision fundamentals, a math and logical reasoning round, sometimes a technical presentation, and hiring-manager or cross-team conversations. One candidate was told interviews are weighed together but was rejected after one weak round. Several candidates say they never heard back after the panel, so follow up and keep other processes moving.
How long does the Zoox ML interview process take?
Reported timelines run Not well reported; one candidate calls the process very long and a take-home review took about 10 days.
What does Zoox ask in ML interviews?
By reported question mix: General coding & language questions 35%; ML & computer vision fundamentals 25%; Math & logical reasoning 15%; Take-home project (DS / CV) 10%; Behavioral & presentation 10%; ML design 5%. The most repeated themes: Self-attention formula: why divide by the square root of d, and why both softmax and scaling; Count the parameters of a CNN layer with 3 x 3 kernels; Parse text records and compute a score with per-type rules.
How should I prepare for the Zoox ML interview?
Expect knowledge checks: memorize core ML and vision formulas and be ready to justify every term. Ask early whether you will code in C++ or Python, and revise that language's details, not just algorithms. Practise quick parameter, shape and compute estimates for convolution and attention layers.
Are there free Zoox interview questions?
Yes. LeetML has 2 questions reported from Zoox interviews, and 1 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- AmazonLeadership Principle questions in every round, standard algorithm coding, and ML breadth, depth and design rounds that differ by track.
- WaymoGraph-heavy coding, numpy ML coding and driving-flavored ML design, plus a research talk for scientists.
- NuroAlgorithm coding, hand-written ML building blocks and a deep dive on your own research, for an autonomous driving company.
- TeslaProject deep dives, practical numpy coding on driving data and layers written from scratch, with a loop shaped by the team you join.
- Applied IntuitionCoding-heavy screens, hands-on transformer debugging and a fast loop, for an autonomy software company that is open about its pace.