Company interview guide
Uber Machine Learning Interview Guide (2026)
Grid and graph coding, ML from scratch and marketplace system design built around ETA, pricing and Uber Eats.
Uber's machine learning interview loop has 8 reported stages, from recruiter call to science track: sql, stats & product cases; typical length: Several weeks.
Roles covered: ML Engineer, Applied Scientist and Data Scientist. The question bank holds 42 questions reported from Uber interviews, 8 of them free to solve in your browser.
- Interview process length
- Several weeks
- Stages
- 8
- Reports linked to bank questions
- 52
- Bank report dates
- Nov 2017–Sep 2026
- Questions in bank
- 42 (8 free)
- Last reviewed
- October 2026
How the Uber interview loop works
A recruiter call, an optional proctored online test (mostly 2021-2024 reports), then one or sometimes two technical screens of about an hour, run on CodeSignal according to Uber's engineering hiring page. The virtual onsite is usually four rounds, sometimes five: algorithm coding, ML coding, ML system design and a hiring manager or leadership round; one design round ran 90 minutes. Science-track loops replace much of the coding with SQL, statistics, experimentation and product cases. Hiring is by team and level can be adjusted after the loop. Several candidates report slow or patchy recruiter communication and round labels that did not match the content, so confirm each round in writing.
The Uber interview process, stage by stage
01Recruiter call
About 30 min call
A short conversation about your background, why Uber and which team or domain you fit. Roles are attached to specific teams, so this call is where you learn whether the opening is about pricing, ETA, Eats ranking, fraud or something else.
What it tests
- Why Uber, and whether you have used the product
- How your past work maps to the team's domain
- Logistics: level, location, timeline
How to prepare
- Ask which team the role sits in and what each later round covers; reported round labels do not always match the content.
- Prepare a two-minute summary of the project most relevant to that team's domain.
Recruiter or talent call as the first step in 21 reports (3 ML-engineer, 18 data and applied scientist), 13 since 2024; the company's engineering, ML and sciences guides all list it, at 30 min.
02Online assessment (some candidates)
Proctored CodeSignal test, 4 problems in about 70-75 min, camera and ID required
Several ML engineer candidates, all in 2021-2024 and including a new-grad pipeline, were sent a timed four-problem test before any live interview. No 2025-2026 ML engineer report mentions one, while two 2026 data scientist reports describe a written test with SQL, so it depends on level, track and location.
What it tests
- Speed on easy to hard array, string and greedy problems
- Choosing which problem to spend time on when the last one carries the most points
- Writing correct code without a human hint
How to prepare
- Budget time by points: bank the easy problems fast and leave the longest block for the last, highest-scoring one.
- Rehearse in the proctored setup (webcam, no copying code) so the environment is not a surprise.
Proctored four-problem online test in 7 ML-engineer reports from 2021-2024, none later; two 2026 data scientist reports describe a written SQL test. The company's ML and sciences guides list no such test.
03Technical screen
About 60 min with a team engineer or manager: intro, ML or project questions, then one coding problem; some candidates had two screens
Most screens pair a short ML or project discussion with one easy to medium algorithm problem, very often on a grid or graph. Some screens instead ask for a derivative-free optimizer, a short ML coding task, an open design chat about where ML fits in a product flow, or a manager-led behavioral hour. An Uber Eats team gave an ML engineer a data-science style screen with pandas and a t-test.
What it tests
- Grid scanning and flood fill: counting islands, distances to blockers, neighbour counts
- Graph modelling of ratios and conversions with BFS or DFS
- Ternary or golden-section search for a black-box unimodal function
- ML basics: logistic regression, calibration, metrics for imbalanced data
- Explaining time and space complexity, then improving space
How to prepare
- Precompute per-direction distances in a grid with one sweep per direction instead of walking from every cell.
- For black-box minimization, state the stopping rule and the evaluation count; golden-section reuses one evaluation per step.
- Be ready to explain calibration and how you would check it with a reliability plot.
- Run a few test cases out loud; some interviewers discourage print debugging.
Free reported questions for this stage
- Why L1 Regularization Produces Sparse SolutionsClassical MLEasySolve
One live screen of coding plus ML or project talk in 16 ML-engineer reports, 11 since 2024; topics vary by team. The company's ML guide lists a 60-min screen on coding plus design and architecture; its engineering page says CodeSignal.
04Onsite: coding
1-2 rounds of about an hour each, live coding
Expect either a classic graph problem or a longer class-design task with several methods that grow in difficulty and finish with a top-K query. Interviewers probe data-structure choices, for example how a heap copes with stale or duplicate entries.
What it tests
- Designing a small stateful class with clean methods
- Maintaining a live top-K ranking under updates
- Aggregating over a tree or referral hierarchy
- Topological ordering and level-by-level scheduling
- Dynamic programming on sequences and counts
How to prepare
- For top-K under updates, decide early between a heap with lazy deletion and a sorted structure, and say why.
- Read long prompts twice and restate every method's contract before coding.
- Parallel task scheduling is Kahn's algorithm counted in waves; also handle a cycle explicitly.
One or two algorithm or class-design rounds in 7 ML-engineer onsite reports, 6 since 2024. The company's ML guide lists a 60-min algorithms and data structures coding interview in the loop.
05Onsite: ML coding
About 60 min, implement a classic algorithm in plain Python or numpy without ML libraries
The ML coding round asks for a textbook algorithm written from scratch and then discussed. Clustering (K-means, or K-medians with L1 distance) is the most reported, followed by linear and logistic regression; one screen asked for a CLIP-style contrastive loss. One candidate's ML coding round was swapped for an ordinary coding round when the interviewer changed.
What it tests
- K-means or K-medians: initialization, assignment, update, convergence
- Linear and logistic regression with gradient descent and L2
- Loss functions written stably, including softmax-based contrastive losses
- Explaining how the algorithm behaves and fails
How to prepare
- Know why the median, not the mean, minimizes L1 distance, and how that changes K-means.
- Write one training loop that serves both linear and logistic regression and explain the gradient difference.
- Practise without sklearn or pandas, and aim to finish K-means in well under 30 minutes; one candidate had that budget.
Free reported questions for this stage
- Linear and Logistic Regression From Scratch With L2ML CodingMediumSolve
From-scratch ML coding (K-means or K-medians, regression, a contrastive loss) in 7 ML-engineer reports, 6 since 2024; one at the screen. The company's ML guide lists a 60-min coding round on depth in specialization.
06Onsite: ML system design
About 60-90 min discussion; one report had ML fundamentals questions mixed in
Prompts come from Uber's marketplace rather than generic recommender templates: real-time ETA from GPS pings, regional surge pricing and Uber Eats home-feed ranking. Interviewers dig into exactly how each feature is computed from the data they allow, and switch to ML fundamentals when you describe the model.
What it tests
- Spatial and temporal features on a hex or grid index, built only from allowed data
- Handling noisy GPS, late events and real-time aggregation
- Pricing control: stability, fraud, demand shocks, supply response
- Ranking with marketplace constraints: ETA, availability, fees, cold start
- Offline metrics, online experiments and monitoring
How to prepare
- Do not assume map or POI data unless the interviewer grants it; explain how you would derive each signal.
- Have a plan for GPS noise (filtering, snapping, outlier rejection) ready; one candidate spent ten minutes on it.
- Explain why marketplace experiments often need switchback or region-level designs rather than user splits.
- Expect follow-ups on tree depth, regularization and dropout when you name your model; one candidate got all three.
Free reported questions for this stage
- Real-Time Remaining ETA per Grid Cell from GPS PingsML System DesignHardSolve
- Regional Surge Pricing System DesignML System DesignHardSolve
- Tree Depth, L1 vs L2 vs Weight Decay, and Dropout at InferenceDeep LearningMediumSolve
ML system design round in 7 ML-engineer onsite reports since 2022 plus 2 science-track reports; prompts are ETA, surge pricing, Eats or feed ranking. The company's ML guide assesses design at the screen, with no separate design round.
07Hiring manager & leadership
About an hour with the hiring manager or a senior engineer, often a project deep dive; some managers run this at the screen stage
A detailed walk through one of your projects followed by standard behavioral questions. The project should match the team's domain: one candidate was told an agent project did not fit a pricing and marketplace team.
What it tests
- Depth and personal ownership on a past ML project, including impact
- Trade-offs you made and how you communicated them
- Conflict, disagreement and cross-functional work
- Mentorship and roadmap influence at senior levels
How to prepare
- Pick a project close to the team's work (pricing, ranking, forecasting, causal inference) and prepare its metrics and failures.
- Prepare one story each for conflict, a hard trade-off and mentoring, with concrete outcomes.
- Know Uber's products and values; recruiters and managers ask whether you have used the app.
Hiring manager, leadership or project deep-dive round in 9 ML-engineer and 8 science-track reports, 14 since 2024; two ran at the screen. The company's ML guide lists a 60-min collaboration and leadership interview.
08Science track: SQL, stats & product cases
Applied and data scientist loops: the sciences guide lists a 30-min recruiter call, a 60-min live coding round, a 60-min Jam Session and 4-6 team sessions of 45 min; a few take-homes
For data scientist and applied scientist roles, and for some ML roles on Uber Eats, the loop centres on SQL or pandas, A/B testing and product cases framed on Uber's own metrics. Most practice problems here come from older community question lists rather than recent first-hand reports, so treat them as topic practice.
What it tests
- SQL joins, window functions, ranking and LAG on time-stamped tables
- Experiment design with spillovers: switchback, region splits, synthetic control, difference in differences
- Product cases on driver incentives, fraud, delivery completion and acquisition cost
- Two-sample tests, variance reduction and sample size
- Probability and expected value with dice-style puzzles
How to prepare
- Be able to explain why a rider-level A/B test leaks in a two-sided marketplace and what to do instead.
- Practise window-function queries without sample rows; one candidate got only column names.
- Structure cases: clarify the goal, propose metrics and guardrails, then hypotheses and data to check.
Free reported questions for this stage
- Alcohol Order Completion Gap and a Two-Sample t-TestA/B Testing & CausalMediumSolve
- Each User's Third PurchaseSQLMediumSolve
SQL, stats, experiment and product-case rounds in 42 data, applied and ML researcher reports (28 since 2024) plus one Uber Eats ML-engineer screen; matches the company's sciences guide.
Questions and themes reported most often
- Proctored four-problem online test in about 70 minutes (7 reports, online assessment (some candidates))
- Product or metric case on an Uber business problem (fraud, delivery, driver incentives) (6 reports, science track: sql, stats & product cases)
- Deep dive on a past ML project, then conflict and trade-off questions (5 reports, hiring manager & leadership)
- SQL with window functions, ranking and LAG on time-stamped trip or order tables (4 reports, science track: sql, stats & product cases)
- Count islands or connected zones in a grid (3 reports, technical screen)
- K-means or K-medians written from scratch (3 reports, onsite: ml coding)
- Design an ETA system: real-time remaining ETA per grid cell or ETA ranges (3 reports, onsite: ml system design)
- Marketplace experiments with spillover: switchback, synthetic control, difference in differences (3 reports, science track: sql, stats & product cases)
What Uber asks about
- Coding (algorithms & class design)32%
- ML / marketplace system design20%
- ML coding & fundamentals17%
- Experimentation, stats & product15%
- Behavioral & project deep dive10%
- SQL & data6%
Uber interview prep checklist
- Solved islands, robots-by-blocker-distance and currency conversion in under 25 minutes each
- Implemented K-medians with L1 distance and regularized logistic regression without libraries
- Can sketch a real-time ETA service from GPS pings, including noise handling
- Can design surge pricing with smoothing, guardrails and a switchback experiment
- Built a top-K leaderboard class that survives repeated updates
- Prepared a domain-matched project deep dive plus conflict and trade-off stories
Tips
- Drill grid and graph problems first: islands variants, per-direction distances and BFS over ratio graphs cover most screens.
- Write K-means, K-medians, and linear and logistic regression from scratch until each takes under 20 minutes.
- Prepare marketplace designs for ETA, surge pricing and Eats ranking, built on a hex grid with real-time features.
- In design rounds, justify every feature from the data you were given; invented signals get challenged.
- Learn why marketplace experiments need switchback or region designs and how to read them.
- Practise long class-design prompts that add methods and end with a top-K query under updates.
- Match your deep-dive project to the team's domain; pricing and marketplace teams want causal, forecasting or ranking work.
- Confirm each round's content with your recruiter, since round labels have not always matched what was asked.
See the rest of the Uber loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Uber loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Uber interview FAQ
How many interview rounds does Uber have for ML roles?
Candidates report 8 stages: Recruiter call, Online assessment (some candidates), Technical screen, Onsite: coding, Onsite: ML coding, Onsite: ML system design, Hiring manager & leadership and Science track: SQL, stats & product cases. A recruiter call, an optional proctored online test (mostly 2021-2024 reports), then one or sometimes two technical screens of about an hour, run on CodeSignal according to Uber's engineering hiring page. The virtual onsite is usually four rounds, sometimes five: algorithm coding, ML coding, ML system design and a hiring manager or leadership round; one design round ran 90 minutes. Science-track loops replace much of the coding with SQL, statistics, experimentation and product cases. Hiring is by team and level can be adjusted after the loop. Several candidates report slow or patchy recruiter communication and round labels that did not match the content, so confirm each round in writing.
How long does the Uber ML interview process take?
Reported timelines run Several weeks (Several weeks; often about one round per week).
What does Uber ask in ML interviews?
By reported question mix: Coding (algorithms & class design) 32%; ML / marketplace system design 20%; ML coding & fundamentals 17%; Experimentation, stats & product 15%; Behavioral & project deep dive 10%; SQL & data 6%. The most repeated themes: Proctored four-problem online test in about 70 minutes; Product or metric case on an Uber business problem (fraud, delivery, driver incentives); Deep dive on a past ML project, then conflict and trade-off questions.
How should I prepare for the Uber ML interview?
Drill grid and graph problems first: islands variants, per-direction distances and BFS over ratio graphs cover most screens. Write K-means, K-medians, and linear and logistic regression from scratch until each takes under 20 minutes. Prepare marketplace designs for ETA, surge pricing and Eats ranking, built on a hex grid with real-time features.
Are there free Uber interview questions?
Yes. LeetML has 42 questions reported from Uber interviews, and 8 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- LyftProbability and marketplace cases for Data Scientists; coding, ML framing and ML design for ML Engineers.
- DoorDashTimed SQL and marketplace case studies for data scientists; project deep dives, coding and infra-heavy ML design for ML engineers.
- InstacartModerate coding, deep questions on classical ML and your own models, and grocery-flavoured ML design and experiment cases.
- AirbnbMarketplace ranking design, a hard look at your past models, one practical coding round and a values check.
- GrabAn online test with ML, SQL and coding, then technical, case-study and manager rounds rooted in ride-hailing, delivery and payments.
- LinkedInProbability-flavoured coding, ML fundamentals and recommender design built around LinkedIn's own products.