Company interview guide
Nubank Machine Learning Interview Guide (2026)
Easy live coding, then track-specific cases: modeling and business for data scientists, architecture and ML programming for ML engineers.
Nubank's machine learning interview loop has 9 reported stages, from talent acquisition call to hiring manager and values interviews; typical length: 3 weeks to 3 months; recent reports usually about 2 months.
Roles covered: Data Scientist and ML Engineer. The question bank holds 5 questions reported from Nubank interviews, 1 of them free to solve in your browser.
- Interview process length
- 3 weeks to 3 months; recent reports usually about 2 months
- Stages
- 9
- Reports linked to bank questions
- 5
- Bank report dates
- Jan 2019–Sep 2024
- Questions in bank
- 5 (1 free)
- Last reviewed
- October 2026
How the Nubank interview loop works
A talent acquisition call, sometimes followed by an online test on a platform, then a live basic programming round of four or five short problems. Data scientists then take a modeling case and a business case; ML engineers take an architecture (ML system design) case and an MLE programming case. The loop ends with a hiring manager or leadership interview and, in many reports, a separate values and behavioral interview with talent acquisition. Earlier interviewers are often peers of the final hiring manager, and in 2026 reports the exact team was only settled in that last interview. Take-home projects were common up to 2022 but appear in only one report since 2023. Parts of the process run in English even in Latin American offices, and gaps between rounds can stretch to weeks.
The Nubank interview process, stage by stage
01Talent acquisition call
About 30 min by video or phone with a recruiter; part of it may be in English
A two-way conversation about your background, the open roles and why you want to join. Some candidates were asked one question in English to check fluency, and a few had an extra early call with a hiring manager or lead data scientist on their past projects. After this call most candidates receive a document describing each later round.
What it tests
- Why Nubank, and how your experience fits the role
- Spoken English for roles where the team works in English
- Logistics: level, location, hybrid availability and timeline
How to prepare
- Prepare a short answer in English about a time you faced a difficulty; it has been asked here.
- Ask for the preparation guide for every round and read it closely; candidates say the instructions are precise.
- Ask which track (data science or ML engineering) you are being assessed for, since one candidate was moved between tracks mid-process.
Opening recruiter or talent call in 23 DS and ML engineer reports plus 3 software engineer reports; English checks in four. The company's data science hiring page lists this talent acquisition interview first.
02Online assessment (some loops)
Timed test on an online platform: short coding tasks plus multiple-choice ML, math or statistics questions
Not every loop has it. Where it appears, it sits between the recruiter call and the live rounds and mixes simple programming with ML theory, math and statistics questions. Data science reports from 2026 describe an online AI and ML test, and one 2025 ML engineer report mentions a logic and financial math challenge.
What it tests
- Basic programming: loops, dictionaries, simple math on integers
- ML theory multiple choice: overfitting, model choice, metrics
- Probability, statistics and basic math under time pressure
How to prepare
- Brush up on quick numeric problems such as counting trailing zeros of a factorial.
- Review core ML definitions so the theory questions go fast and leave time for the code.
An unproctored online test in 6 DS reports from 2019 to 2026, two of them in 2026; ML engineer reports rarely mention it. The company's hiring page does not list it.
03Basic Programming Case (live coding)
About 60 min, screen shared on an online coding platform with one or two engineers; 4-5 short problems in Python
The most consistent round across both tracks. Expect four or five warm-up to easy problems on arrays, dictionaries, strings and stacks, solved while talking through your reasoning. Several candidates finished early; interviewers care about clear reasoning and collaboration more than speed. A few older data science loops added SQL or pseudo-code here.
What it tests
- Clean, correct Python for small array, string and dictionary tasks
- Using a stack, heap or sort where it fits
- Thinking aloud and taking hints from the interviewer
- Basic complexity awareness (not always asked)
How to prepare
- Practise easy problems at a steady pace; aim to finish four in about 40 minutes.
- Narrate your plan before coding; reports say collaboration is weighed heavily.
- Data science candidates should also be ready for a simple SQL query in case the team includes one.
Live easy-level coding in 25 DS and ML engineer reports from 2019 to 2026, usually four or five problems in about an hour. The company's hiring page lists basic programming as the first technical interview for both tracks.
04Take-home case (older and some DS loops)
Usually 3 days to a week: build a tabular classification model with a short report, later reviewed in an interview
Up to 2022 most data science loops included a take-home, such as a binary classification pipeline, a churn model or a write-up of how you would model a business phenomenon, and a later round walked through your choices. Since 2023 only one data science report mentions it, so treat it as possible but not expected.
What it tests
- Feature engineering and evaluation on tabular data
- Choosing and justifying a metric when none is given
- Writing a clear, short report of decisions and results
- Defending your choices in a follow-up review
How to prepare
- If the brief has no metric, state the one you chose and why at the top of the report.
- Keep the code readable; reviewers question it line by line in the follow-up round.
Take-home in 9 DS reports, 8 of them from 2019 to 2022 and one churn project in 2024; no ML engineer report since 2023 mentions one.
05Data science: Modeling Case
About 60 min discussion with a data scientist, built around a modeling problem from the company's context
A conversation on how you would structure an ML solution: data cleaning, missing values, categorical encoding, feature selection, model choice, validation and calibration. Questions are about intuition more than math, and often start from a credit or pricing scenario. Some reports describe it as a rapid run of short technical questions.
What it tests
- Imputation strategies and how they affect different models
- Random forest versus gradient boosting, overfitting and bias-variance
- Encoding categorical variables, normalization versus standardization
- Probability calibration: why, how and how to check it
- Feature selection, PCA trade-offs and explaining models with SHAP
How to prepare
- Have a crisp end-to-end story for a tabular credit model, from raw data to monitoring.
- Be ready to argue why PCA can hurt an interpretable credit model.
- Answer short questions briefly, then offer depth; some interviewers move fast.
Modeling or ML concepts round in 16 DS reports from 2019 to 2026, with imputation, tree ensembles and calibration recurring. The company's hiring page lists modeling as a DS technical interview.
06Data science: Business Case
About 60 min, a realistic problem discussed live with a data scientist
Often called the hardest data science round. You get a real situation, such as a key metric or a production model whose performance dropped, and must investigate causes and propose a technical answer. Older loops framed it as applying ML to a business problem such as credit approval.
What it tests
- Structured diagnosis of a metric or model performance drop
- Spotting data drift, label changes and pipeline bugs
- Turning a business question into an ML or analytics solution
- Prioritizing hypotheses and proposing next steps
How to prepare
- Use a fixed checklist: data pipeline, population shift, label definition, seasonality, product changes.
- Tie every hypothesis to a check you would run and what result would confirm it.
Free reported questions for this stage
- Detecting and Handling Data Shift in Deployed ModelsML System DesignMediumSolve
Business case in 8 DS reports from 2019 to 2025; three describe investigating a dropped metric or model. The company's hiring page lists business analysis as a DS technical interview.
07ML engineering: Architecture Case
About 60 min, whiteboard-style ML system design with no code
You are given a business problem, for example reducing messages that reach customer support, validating loans or running batch predictions, and guided through designing an ML solution. Expect questions on the whole lifecycle: framing, datasets, model choice, serving, metrics, monitoring and fixing bad predictions. Interviewers want to hear trade-offs argued, not one right answer.
What it tests
- Framing a fintech business problem as an ML task
- Batch versus real-time serving and the services around a model
- Choosing offline and online metrics and monitoring drift
- Feedback loops for correcting wrong predictions
- Cloud data architecture choices for storage and processing
How to prepare
- Agree on scope and the success metric in the first minutes.
- Walk the ML lifecycle in order so you do not skip monitoring or retraining.
- Prepare one design for a support-ticket or customer-contact problem; it has recurred.
Architecture or ML system design case in 10 ML engineer reports from 2022 to 2026, consistently discussion only. The company's hiring page lists architecture as an ML engineer technical interview.
08ML engineering: MLE Programming Case
About 60 min of pair programming on a small prepared ML codebase; about 4 questions
You work in an existing project, for example code that loads a trained model and produces outputs, or an ML experiment workflow. Questions ask you to understand the code, find problems in the current solution and add simple features; unit tests are optional. Some candidates found the statement long and could not see later questions in advance. A 2026 report built the round around analyzing an experiment run across app screens.
What it tests
- Reading and extending unfamiliar code quickly
- Using a trained model correctly at a business level
- Spotting flaws in an existing ML workflow
- Basic statistics for experiments and model outputs
How to prepare
- Spend the first minutes asking questions until the problem is clear.
- Write simple, readable code; finishing all four questions is not required.
- Review A/B test analysis basics, since an experiment scenario has appeared.
ML programming on existing code in 6 ML engineer reports from 2022 to 2026; a 2026 DS report also had an AI live-coding round. The company's hiring page lists advanced programming for ML engineers.
09Hiring manager and values interviews
1-2 conversations of about 60 min: the hiring manager or leadership, and often a separate values and behavioral interview with talent acquisition
The leadership interview checks seniority and values and explains the team's challenges; candidates say it carries heavy weight in the decision. It covers past projects, prioritization, stakeholders and how you would redo a project that failed. Many loops add a behavioral interview on the company's values, and interviewers debrief together to decide fit and level.
What it tests
- Depth and impact of your past projects
- Prioritization and stakeholder management
- Fit with the company's values and self-reflection
- Seniority signals: ownership, delegation, handling feedback
How to prepare
- Read the company's values and prepare one story per value.
- Prepare a failed project and how you would approach it today.
- Have a sharp answer to why Nubank; some interviewers ask if you are a customer.
Hiring manager or leadership round in 19 DS and ML engineer reports plus 3 software engineer reports; a separate values interview in nine. The company's hiring page ends with a leadership interview on values and seniority.
Questions and themes reported most often
- Four or five easy array, string and stack problems in an hour (9 reports, basic programming case (live coding))
- Why Nubank, and which of its values you identify with (7 reports, hiring manager and values interviews)
- Design an ML solution for a fintech business problem, such as cutting customer-support contacts (6 reports, ml engineering: architecture case)
- Extend a small existing ML codebase in a pair-programming session (6 reports, ml engineering: mle programming case)
- Overfitting, bias-variance and random forest versus gradient boosting (5 reports, data science: modeling case)
- Features and design for a credit card or loan approval model (4 reports, data science: business case)
- A failed or challenging past project and what you would change (4 reports, hiring manager and values interviews)
- How to impute missing data and why it affects models differently (3 reports, data science: modeling case)
What Nubank asks about
- General coding28%
- ML modeling & concepts22%
- ML system design / architecture17%
- Business & product cases13%
- Behavioral, values & manager14%
- ML programming on existing code6%
Nubank interview prep checklist
- Solved four easy coding problems in Python in under 45 minutes while thinking aloud
- Can explain imputation choices, random forest versus boosting and calibration in plain words
- Practised a metric-drop investigation from hypothesis to check
- Sketched an ML system for a fintech problem from framing to monitoring
- Extended an unfamiliar small codebase within an hour
- Prepared value-based stories, including a failed project, in English
- Read the preparation guide for every round
Tips
- Treat the basic programming round as a fluency check: practise easy problems out loud until four take about 40 minutes.
- Pick your track early and drill its two cases: modeling and business for data science, architecture and MLE programming for ML engineering.
- Build one full credit-model story covering features, imputation, class imbalance, calibration and monitoring.
- Practise diagnosing a metric drop with a fixed checklist of data, population, labels and product changes.
- For architecture, walk the ML lifecycle in order and argue trade-offs; no code is expected.
- Read each round's preparation guide; candidates say it describes the round accurately.
- Prepare behavioral answers in English, since parts of the loop often run in English.
- Plan for a long process with gaps between rounds and follow up with your recruiter if things stall.
See the rest of the Nubank loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Nubank loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Nubank interview FAQ
How many interview rounds does Nubank have for ML roles?
Candidates report 9 stages: Talent acquisition call, Online assessment (some loops), Basic Programming Case (live coding), Take-home case (older and some DS loops), Data science: Modeling Case, Data science: Business Case, ML engineering: Architecture Case, ML engineering: MLE Programming Case and Hiring manager and values interviews. A talent acquisition call, sometimes followed by an online test on a platform, then a live basic programming round of four or five short problems. Data scientists then take a modeling case and a business case; ML engineers take an architecture (ML system design) case and an MLE programming case. The loop ends with a hiring manager or leadership interview and, in many reports, a separate values and behavioral interview with talent acquisition. Earlier interviewers are often peers of the final hiring manager, and in 2026 reports the exact team was only settled in that last interview. Take-home projects were common up to 2022 but appear in only one report since 2023. Parts of the process run in English even in Latin American offices, and gaps between rounds can stretch to weeks.
How long does the Nubank ML interview process take?
Reported timelines run 3 weeks to 3 months; recent reports usually about 2 months.
What does Nubank ask in ML interviews?
By reported question mix: General coding 28%; ML modeling & concepts 22%; ML system design / architecture 17%; Business & product cases 13%; Behavioral, values & manager 14%; ML programming on existing code 6%. The most repeated themes: Four or five easy array, string and stack problems in an hour; Why Nubank, and which of its values you identify with; Design an ML solution for a fintech business problem, such as cutting customer-support contacts.
How should I prepare for the Nubank ML interview?
Treat the basic programming round as a fluency check: practise easy problems out loud until four take about 40 minutes. Pick your track early and drill its two cases: modeling and business for data science, architecture and MLE programming for ML engineering. Build one full credit-model story covering features, imputation, class imbalance, calibration and monitoring.
Are there free Nubank interview questions?
Yes. LeetML has 5 questions reported from Nubank interviews, and 1 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- Capital OneA timed four-question coding assessment, then a four-round Power Day of ML technical, a business case, behavioral and, on recent ML engineer loops, ML system design.
- StripePractical multi-part coding, a timed dataset-to-model round, an ML bug squash and payments-flavoured ML design.
- PayPalSQL and Python screens, medium-difficulty algorithm coding, ML fundamentals and fraud-flavoured cases, with loops that vary by team.
- RobinhoodSQL on trading data, probability and churn modeling for data scientists; coding, ML design and a project deep dive for ML engineers.
- Block (Square)Live pairing on practical Python and SQL, then a long onsite of hands-on modeling, case studies and ML Q&A.
- CoinbaseCognitive and culture tests up front, hard SQL, live modeling on messy data, and a work-trial presentation for data scientists.