Uber Machine Learning Engineer Interview Process

Uber's Machine Learning Engineer loop filters hardest before the onsite. Two or three phone screens mix a coding problem, an ML discussion and a deep dive into your own project, and the coding is often numerical rather than catalog material: minimizing a black-box convex function, placing pickup hubs under Manhattan distance. At senior and Staff level the project conversation sets the outcome, and fit with the hiring team's domain (pricing, marketplace, recommendations or ML infrastructure) shapes how even strong work is read.

This guide breaks down each stage of the Uber Machine Learning Engineer interview, what interviewers evaluate in each round, and how to prepare, with questions candidates were actually asked.

Verified: Based on 8 candidate-reported interview experiences, primarily from 2025 Q3 through 2026 Q2, cross-checked against recent public accounts.

Uber Machine Learning Engineer interview process

Uber hires Machine Learning Engineers into specific teams, and the loop bends toward that team: interviewers adapt sessions to the strength a candidate states, and modeling teams that ship into the marketplace add an experimentation round. The usual shape is a recruiter call, two or three phone screens of 45 to 60 minutes, and a final round of one to several sessions, typically over three to six weeks.

Mid-level loops weigh coding and ML system design; from L5A (senior) upward the project and leadership conversations decide the level, and Staff loops can be almost entirely conversation.

Here's what the interview process can look like:

• Recruiter screen: Background, target level and which ML surface you work on; the answer steers later rounds.
• Technical phone screen (one or two): 45 to 60 minutes of coding, often a numerical problem, sometimes followed by an ML systems discussion.
• Project deep dive (phone or onsite): One past ML project examined for formulation, metrics, impact and what you would change.
• ML coding interview: A small probabilistic model built from scratch in plain Python.
• ML system design interview: About 60 minutes on a recommender, an ML pipeline or a data platform that feeds models.
• Experimentation interview (marketplace and applied-science loops): Experiment design, switchbacks and confidence intervals.
• Leadership and behavioral interview: Mentoring, roadmaps, cross-functional work and disagreements.

Tip: State your strongest area in one sentence on the recruiter call and at the start of each screen. A candidate who named inference optimization had the screen tailored to serving and latency, while one whose generative-AI work did not match a pricing team's domain was told a downlevel was likely.

Technical phone screen

Coding screens run 45 to 60 minutes in a shared editor, often in Python. For senior candidates the recurring problem is numerical: two separate reports describe minimizing a convex function on an interval that can only be evaluated at points, with no closed form or derivative. Ternary search is the expected first answer; golden-section search, which reuses one evaluation per iteration, is the improvement interviewers probe for, along with stopping rules for continuous versus discrete domains.

Other screens use applied algorithm problems: choosing K pickup hubs to minimize total Manhattan distance (which separates into independent one-dimensional problems), the minimum time to finish unit-length tasks under precedence constraints (a longest path after topological sort), squares of a sorted array with negatives, and reversing a linked list with follow-ups.

Interviewers evaluate:

• Recognizing the structure: Seeing unimodality, L1 separability or a critical path instead of reaching for brute force.
• Cost reasoning: Function-evaluation counts, convergence rate, and time and space bounds.
• Termination and precision: When to stop narrowing an interval and how tolerance is defined.
• Improving the first version: Proposing the better method without being led to it.

Practice ternary and golden-section search, binary search on real values and topological sort in Python until each takes under 15 minutes.

Note: The working answer is only the start. One senior candidate implemented ternary search, was asked to cut the number of evaluations, ran out of time before coding golden-section search, and was rejected.

Sample questions for the technical phone screen

Here are some questions reported by recent candidates:

• Minimize a convex black-box function on an interval to a given tolerance, then reduce the number of evaluations.

• Choose exactly K hub locations to minimize the sum of Manhattan distances from each stop to its nearest hub.

• Given unit-time tasks and precedence constraints, find the minimum time to finish all of them.

• Strictly Convex Function Minimum

• K Shuttle Pickup Locations

• Task Dependency Completion Time

• Squares of a Sorted Array

Project deep dive

Most loops examine one past ML project, often in the first screen and sometimes as the opening half of an onsite session. Expect up to an hour on formulation, data, model choice, evaluation metrics, measurable business impact and team structure, sometimes extended into system architecture, latency and orchestration.

At Staff level the conversation adds a harder question: starting over, how would you redefine the core problem from first principles? A Staff candidate who answered by improving the existing approach was rejected because the depth of ML insight was judged below target. Domain fit is judged here too; a senior candidate's generative-AI agents project read as a mismatch for a pricing and growth team.

Interviewers evaluate:

• Problem formulation: The real objective, the framings you rejected, and why the chosen one fit the business.
• Measurable impact: How success was defined and which metric moved, by how much.
• Individual ownership: Your own architecture and modeling decisions, separated from the team's.
• Reframing at senior levels: A different approach you would take now, not an increment on the current one.

Prepare two projects with a written counterfactual for each: the objective restated from scratch, two alternative framings, and what you would build today.

Sample questions for the project deep dive

Here are some questions reported by recent candidates:

• Walk through a significant ML project: formulation, data handling, model choice, evaluation and what you would improve.

• How would you redefine this project's core problem from fundamentals rather than iterate on the current solution?

• Explain the architecture of a multi-project system you designed, including latency limits and orchestration.

• Describe an ML system you built

• Describe your strongest project

• Demonstrate business impact from a project

ML coding interview

ML coding asks for a small model or sampler from scratch, without libraries doing the core work. The reported onsite version followed a project deep dive in the same session and had two parts: map each word in a corpus to the words that follow it with their counts, then generate text from a start word by weighted random choice until an end token or a length limit.

Interviewers evaluate:

• Data structure choice: A transition map that is easy to update and sample from, justified out loud.
• Sampling correctness: Probabilities proportional to observed counts, with dead ends and end tokens handled.
• Execution speed: Both parts working and checked on a toy corpus.

Practice weighted sampling, n-gram counts and simple regression in plain Python, 20 minutes each.

Sample questions for the ML coding interview

Here are some questions reported by recent candidates:

• Build a word-transition frequency map from a list of sentences, then a generator that samples each next word by observed frequency.

• ML Coding from Scratch (Regression / Markov / Facility)

• Implement weighted sampling without replacement

• Convert a PDF to a CDF

ML system design interview

This round appears as a phone screen or a 60-minute onsite session in two forms. The product form starts from an Uber goal, such as restaurant recommendations for Uber Eats: data pipelines, features (order history, cuisine, location, time of day), two-tower retrieval, cold start, precision and recall at k, online engagement metrics and scaling. The platform form is broader; one onsite prompt asked for a system to analyze user behavior data from the Uber app with no further constraints, and covered event instrumentation, batch and streaming layers, features for churn and ETA models, schema evolution and session-based funnels.

Infrastructure-leaning screens go into serving instead: model choice under latency budgets, distillation, quantization, dynamic batching, versioning, A/B rollout and drift monitoring.

Interviewers evaluate:

• Scoping a vague prompt: Users, scale, latency and consumers settled before components.
• The full ML path: Data, features, model, offline and online evaluation, and post-launch monitoring, none skipped.
• Explicit trade-offs: Latency against consistency and cost against freshness, defended when pushed.
• Data quality: Out-of-order and duplicate events, schema changes and replaying history.

Practice one ranking system and one event pipeline that feeds model features, 45 minutes each.

Tip: Raise cold start, data quality and monitoring before the interviewer does. In the data-platform round the interviewer kept pushing on late events, duplicates and freshness, and the candidate called it a data-engineering deep dive rather than a product design.

Sample questions for the ML system design interview

Here are some questions reported by recent candidates:

• Design a restaurant recommendation system for Uber Eats, including cold start and offline and online evaluation.

• Design a system to analyze rider behavior events for daily analytics and real-time dashboards, then support funnel analysis.

• ML System Design (Recommendation / Feed Ranking / ETA)

• Design A Nearby Restaurant Recommendation System

• App User Behavior Data Platform

Experimentation interview

Marketplace-facing and applied-science loops can spend one or two rounds on measurement. One reported loop paired a linked-list coding round with a round on how ETA affects rider session conversion, including switchback designs, and a third on launch guardrails, treatment buckets with multipliers and correct reading of confidence intervals.

Interviewers evaluate:

• Marketplace-aware design: Knowing when rider-driver interference calls for a switchback.
• Ratio metrics: Handling conversion-style ratios, including the variance of the denominator.
• Launch safety: Guardrail metrics and staged rollouts.
• Statistical interpretation: What a confidence interval does and does not say.

Prepare five-minute answers on a switchback design and a ratio-metric analysis.

Note: A friendly round is not a pass. That candidate described enthusiastic interviewers and was still rejected, concluding that prompt, precise answers on design principles and economic metrics were what counted.

Sample questions for the experimentation interview

Here are some questions reported by recent candidates:

• Design an experiment to measure how ETA affects rider session conversion.

• Set up a switchback experiment and choose guardrails for the production launch.

• Configure treatment buckets with multipliers, then interpret the resulting confidence intervals.

• Scientist Onsite: Experiment Design (Switchback + CI Interpretation)

• Evaluate ETA Impact on Conversion

• Design a switchback and choose block length

Leadership and behavioral interview

This round typically runs 45 to 60 minutes. It opens with an ML project you led (architecture decisions and trade-offs), then covers mentoring junior engineers, working with product and engineering partners, setting technical roadmaps and resolving disagreements. At Staff level the focus is how you influenced technical direction across teams, not only within your own scope.

Interviewers evaluate:

• Scope matching the level: Team-level delivery for L5A, cross-team direction for Staff.
• Mentoring with outcomes: Who you developed and what changed for them.
• Handling disagreement: The evidence used and how the decision was settled.

Prepare six stories with metrics, including a roadmap you set and a disagreement you resolved with evidence.

Sample questions for the leadership and behavioral interview

Here are some questions reported by recent candidates:

• How have you mentored junior engineers, and what changed as a result?

• How did you set a technical roadmap with product and engineering partners?

• Tell me about a disagreement within your team and how it was resolved.

• Onsite BQ + Hiring Manager + Project Deep-Dive

• Answer leadership questions on tradeoffs and collaboration

• How do you manage performance and disagreements?

How Uber levels and evaluates Machine Learning Engineer candidates

The target level is set early and re-tested throughout, mostly in the project and leadership conversations. Downleveling is a live outcome: a candidate targeting L5A heard after two phone screens that L4 was more likely, and withdrew. Domain fit with the hiring team carries weight alongside technical performance, and a Staff rejection cited depth of ML insight rather than any single failed answer. A rejection triggered a six-month cooldown before applying to other Uber roles, and detailed feedback is rare.

How to prepare for the Uber Machine Learning Engineer interview

1. Drill numerical search: Implement ternary and golden-section search with explicit tolerances and explain evaluation counts.
2. Write a counterfactual for each project: The objective from scratch, two alternative framings, and what you would build today.
3. Match your stories to the team: Learn whether it owns pricing, ETA, recommendations or ML infrastructure, and lead with the closest work.
4. Code small models from scratch: Weighted sampling, n-gram maps and simple regression, finished and tested in 20 minutes.
5. Rehearse two Uber-shaped designs: A ranking system with cold start and a behavior-event pipeline feeding features, each with monitoring.
6. Prepare serving specifics: Distillation, quantization, batching and drift detection, with numbers from your own systems.
7. Know marketplace experiments: Switchbacks, ratio metrics, guardrails and confidence intervals, answered quickly.

About the Uber Machine Learning Engineer role

Uber Machine Learning Engineers build models that make live marketplace decisions (pricing, ETAs, dispatch, recommendations) and own them through serving, latency budgets and monitoring, not only training. Product teams build on Michelangelo, Uber's internal platform for features, training and real-time serving, and platform teams build Michelangelo itself. Corporate staff have worked in the office Tuesday through Thursday since June 2025.

Uber hires Machine Learning Engineers across product and platform teams. Teams include:

• Marketplace pricing and incentives: Real-time rider and driver pricing, incentive simulation and optimization.
• Marketplace signals: ETA prediction, supply availability and demand forecasting.
• Delivery: Pricing and ranking across food, grocery and other delivery verticals.
• AI platform (Michelangelo): Distributed training, feature computation and real-time model and generative-AI serving.
• Uber Freight: Pricing, matching and recommendation for freight.

Machine Learning Engineer experience and education requirements

Postings ask for a bachelor's degree in computer science, engineering, mathematics or a related field, with a master's or PhD often preferred; Staff roles typically ask for six or more years of production ML. Python is standard, with Java or Go on platform teams, and PyTorch is commonly preferred. Senior postings add large-scale data pipelines and Spark or Ray.

FAQs about the Uber Machine Learning Engineer interview

How long does the Uber Machine Learning Engineer interview process take?

Typically three to six weeks from recruiter call to decision, longer when screens are spread out or team alignment takes time.

What is the hardest round of the Uber Machine Learning Engineer interview?

For senior and Staff candidates, the project deep dive, because the reframing question arrives inside what feels like a friendly retrospective. For others, the numerical-optimization screen is the most common surprise.

Is there an online assessment?

None of the Machine Learning Engineer reports included one; the first technical step was a live phone screen in a shared editor.

What happens after a rejection?

Candidates reported a six-month cooldown before applying to other Uber roles, and little detailed feedback.

---

Other Tracks at Uber

• Uber Software Engineer Interview Process