Company interview guide
Capital One Machine Learning Interview Guide (2026)
A 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.
Capital One's machine learning interview loop has 9 reported stages, from recruiter screen to power day: behavioral & job fit; typical length: 1-3 months.
Roles covered: ML Engineer, AI Engineer, Data Scientist and Applied Scientist. The question bank holds 31 questions reported from Capital One interviews, 3 of them free to solve in your browser.
- Interview process length
- 1-3 months
- Stages
- 9
- Reports linked to bank questions
- 31
- Bank report dates
- Nov 2019–Sep 2026
- Questions in bank
- 31 (3 free)
- Last reviewed
- October 2026
How the Capital One interview loop works
A recruiter call, then a 70-minute online coding assessment of four questions for ML and AI engineers (data scientists get a 90-minute pandas and modelling assessment or a week-long take-home), and for many candidates a 30-60 minute hiring-manager screen, before or after the assessment. Several recent ML engineer reports went straight from the assessment to the final stage. The final stage is a Power Day of four one-hour video interviews on one day: for ML engineers usually ML technical, ML system design, a case and behavioral; for data scientists technical, case, a statistics role-play and job fit. Rounds are scored separately and combined. Team matching can follow the Power Day and add weeks.
The Capital One interview process, stage by stage
01Recruiter screen
About 30 min phone or video call
A conversation about your resume, why Capital One and the role. The company's hiring page also lists a short automated assessment of work style after you apply, though candidates rarely mention it.
What it tests
- A clear summary of your ML experience and what you shipped
- Why Capital One and why this role
- Logistics: location, level, timeline
How to prepare
- Ask which assessment you will get (coding, data science or a take-home) and how long you have to finish it.
- Ask whether you are interviewing for a specific team or for a pool that matches teams later.
A recruiter or HR call opens the loop in 36 candidate reports across ML engineer, AI engineer and data scientist roles, 2018-2026; the company's hiring page lists it too.
02Online coding assessment
70 min, 4 questions, proctored with camera and screen share; usually 5-14 days to start it
The most consistently reported step for ML and AI engineers. Two quick warm-ups with long statements, then a tedious simulation or matrix task, then a harder counting or string problem. Questions reward careful implementation more than clever algorithms. One 2026 report describes a longer multi-part version that also covered ML concepts, design and statistics.
What it tests
- Reading long problem statements quickly and coding them without mistakes
- Grid and matrix manipulation: rotations, right/down path DP, piece placement
- Interval counting with a sweep or difference array
- Hash-map counting over transformed values (digit swaps, pairs)
- Passing hidden tests within time limits
How to prepare
- Bank the two easy problems in the first 15 minutes, then spend most of the time on the third.
- The simulation question is long to type and copy and paste may be blocked; write small helpers for bounds checks and placement.
- Watch for timeouts on the last problem; candidates report that even near-full scores do not guarantee a next step.
Four coding questions in about 70 minutes in 18 ML and AI engineer reports, 2021-2026, with the same easy-easy-medium-hard shape; the hiring page mentions an hour-long role test.
03Data science assessment (data scientist track)
90-min timed pandas and modelling assessment (most 2026 reports) or a take-home of about a week; discussed again on Power Day
Data scientist candidates either take a timed assessment of data cleaning and pandas questions ending with training a model, or build a predictive model on a provided dataset and write up the results. Recent reports lean toward the timed version; a credit-card fraud take-home of several hundred thousand transactions has come up more than once. ML engineer reports rarely mention this step.
What it tests
- Data cleaning: missing values, sentinel codes, leakage
- Feature engineering from transaction histories
- Modelling a rare positive class and choosing a fair metric
- A clear written explanation of results and limits
How to prepare
- Answer every question in the prompt explicitly; one candidate who wandered off the brief reports failing here.
- For the timed version, practise pandas cleaning, group-bys and fitting a scikit-learn or XGBoost model quickly, and check hidden test cases.
- Use time-based splits for transaction data and explain why.
- Expect the Power Day to probe your choices, so keep notes on what you tried.
A timed data assessment or a take-home data challenge appears in 30 data scientist reports; 2026 reports describe a 90-minute timed test, earlier ones a week-long take-home. The hiring page lists the challenge.
04Hiring manager screen
30-60 min video call, before or after the assessment; skipped in some ML engineer loops
Mostly a deep dive into your recent projects with behavioral questions woven in. On LLM-focused teams it turns technical, with questions on attention, fine-tuning and inference optimization, sometimes at the level of the math, and one report had live ML coding in a notebook. It is near universal for data scientists; several recent ML engineer reports went from the assessment straight to the Power Day.
What it tests
- Ownership, scope and impact of your current project
- A technical challenge you solved and how you debugged it
- Cross-team conflict, for example an infrastructure bug that hit your service
- LLM fundamentals: self-attention, LoRA, quantization, retrieval-augmented generation
How to prepare
- Prepare one project you can explain end to end, including what you would change and how big its scope was.
- For LLM teams, be ready to write out how LoRA modifies a weight matrix and why it saves memory.
- Leave time for your own questions; these calls often run over.
Free reported questions for this stage
- How Self-Attention Works, From Intuition to Tensor ShapesLLMs & GenAIMediumSolve
A hiring-manager screen appears in 25 reports, mostly data scientist ones; its order relative to the assessment varies, and some ML engineer loops skip it. The hiring page lists a 30-minute pre-screen.
05Power Day: ML technical & job fit
1-2 x 60 min; one round can be about 30 min of coding followed by ML questions
A fast tour of ML fundamentals tied to examples, plus how you would deploy and monitor a model. ML engineer loops usually include a short coding problem, such as an easy tree problem or a prime sieve, inside one of these rounds. In one detailed report the second technical round paired the tree problem with productionizing a fraud model that missed its latency budget. Data scientist versions focus on model comparison, the ML lifecycle and your take-home.
What it tests
- Bias-variance, over- and underfitting, regularization
- Parameters vs hyperparameters, bagging vs boosting, tree ensembles
- Batch vs real-time inference and training on data too big for one machine
- Deployment, rollback and monitoring of drift and business metrics
- Cutting p99 latency of an online scoring API without losing accuracy
How to prepare
- Answer each concept with a concrete example from your own work; interviewers move quickly.
- Have a ready walk-through of the full ML lifecycle on one project, from problem framing to monitoring.
- For latency, first say how you would measure where time goes, then cover features, model size and serving.
Free reported questions for this stage
- ML Fundamentals Rapid-Fire: Bias, Serving and ScaleClassical MLMediumSolve
- Productionizing a Fraud Model: Rollout, Monitoring and a 50 ms p99 BudgetML System DesignHardSolve
One or two technical or job-fit rounds on ML fundamentals, deployment and projects in 25 reports across ML engineer and data scientist roles; the hiring page lists one or two job-fit interviews.
06Power Day: ML system design
About 60 min, verbal or whiteboard; mainly on recent ML engineer loops
Two 2026 ML engineer reports list a separate design round, and two data scientist reports mention ML or infrastructure design, but candidates have shared no concrete prompts. Older ML engineer loops had two ML technical rounds instead. Expect a banking-flavoured system such as real-time fraud scoring, where latency, class imbalance, monitoring and explainability all matter.
What it tests
- Scoping a vague business goal into an ML system
- Online feature computation and serving under a tight latency budget
- Handling rare positives, delayed labels and drift
- Explainability and risk controls expected in a regulated bank
How to prepare
- Ask clarifying questions first and state trade-offs as you go.
- Cover retraining and monitoring, not just the first model.
Named as a Power Day round in 2 ML engineer reports from 2026 and 2 data scientist reports; no prompt details. The fraud-scoring framing comes from public write-ups, not first-hand reports.
07Power Day: case interview
60 min, interviewer-led business problem, sometimes ending in code
A practical business problem close to the role. The ML engineer version reported most recently is a virtual credit card case: weigh the product for customers and the bank, turn written rules into transaction checks, then debug provided code against those rules. Older ML engineer cases covered cloud compute cost, a cost-benefit of software testing and connecting to an acquired bank. Data scientist cases use operations data such as airline schedules or call-center staffing.
What it tests
- Product reasoning about a Capital One style card or banking feature
- Turning precise written business rules into correct logic
- Debugging someone else's code carefully against a spec
- Quantitative trade-offs, for example cost versus benefit of a process change
How to prepare
- Read Capital One's current card and AI products before the day; cases often reference them.
- Take notes on every rule and test each branch; the debugging step is not hard but punishes skipped details.
- Talk through your reasoning and ask for missing numbers instead of guessing.
A case round appears in 26 reports across ML engineer and data scientist roles; topics vary from product and cost trade-offs to operations data. The hiring page lists a case interview.
08Power Day: statistics role-play (data scientist track)
60 min; often about 15 min to prepare from slides, then about 15 min to present, then questions
Data scientist Power Days include a statistics round, usually a role-play in which you explain a model or regression output to a non-technical client, for example flight-delay data with a choice of logistic regression or a decision tree. Some candidates instead got a statistics discussion on a sample dataset and model.
What it tests
- Interpreting regression and tree model outputs correctly
- Explaining results in plain business terms
- Judging whether a given model is appropriate and what would be better
- Answering follow-up statistics questions under time pressure
How to prepare
- Practise explaining logistic regression and decision tree outputs in two minutes each, without jargon.
- Use the preparation window to pick two or three takeaways and a recommendation, not to cover every chart.
A statistics round, mostly a role-play presentation to a non-technical audience, in 12 data scientist reports from 2023-2026; not described for ML engineers or on the hiring page.
09Power Day: behavioral & job fit
45-60 min structured interview
A structured round where the interviewer reads set questions, takes notes and often summarizes your answer back to check it. Some loops fold job-fit questions about your background and motivation into this hour.
What it tests
- Challenging the status quo
- Delivering on time despite unexpected problems
- Prioritizing and working under pressure
- Explaining technical results in plain business terms
How to prepare
- Prepare three or four stories with measurable results, and let the interviewer finish summarizing before adding detail.
- Keep each story short; interviewers tend to move through several prompts in the hour.
A behavioral or job-fit round with tell-me-about-a-time prompts in 24 reports across roles, consistently described as scripted; the hiring page lists it.
Questions and themes reported most often
- Four-question timed assessment: two warm-ups, a tedious simulation, a harder counting problem (14 reports, online coding assessment)
- Resume and project deep dive with the hiring manager (10 reports, hiring manager screen)
- ML fundamentals rapid-fire: bias-variance, overfitting, boosting vs bagging (6 reports, power day: ml technical & job fit)
- Explain a model or regression output to a non-technical audience (6 reports, power day: statistics role-play (data scientist track))
- Walk through the full ML lifecycle, then one end-to-end project (5 reports, power day: ml technical & job fit)
- Transformers, self-attention and LoRA on LLM-focused teams (5 reports, hiring manager screen)
- Business case on operations data: airline schedules, flight delays, call-center staffing (5 reports, power day: case interview)
- Review or debug someone else's code (4 reports, power day: case interview)
What Capital One asks about
- General coding28%
- ML concepts & fundamentals22%
- Business case15%
- Behavioral & job fit12%
- ML system design & deployment10%
- Statistics & model interpretation8%
- LLM concepts5%
Capital One interview prep checklist
- Finished a four-problem, 70-minute timed set with the first two done in under 15 minutes
- Solved a right/down grid DP and a lamp-coverage sweep without hints
- Can explain bias-variance, boosting vs bagging and batch vs real-time inference with your own examples
- Can walk through deploying, monitoring and meeting a 50 ms latency budget for a fraud model
- Practised turning written card-transaction rules into code and debugging it against tests
- Prepared three behavioral stories with measurable results, including one about challenging the status quo
Tips
- Treat the coding assessment as the main filter for ML and AI engineers: practise four mixed problems in 70 minutes with no autocomplete and no copy and paste.
- Drill implementation-heavy problems (grid paths, matrix rotation, placement simulations, interval sweeps) more than advanced algorithms.
- Prepare crisp, example-backed answers to the classic ML fundamentals; the technical rounds cover many topics quickly.
- Know how you would deploy, monitor and speed up a fraud model, since card fraud is the default example in several rounds.
- Study Capital One's card products and public AI work before the case interview.
- In the case and debugging steps, write down every rule and check each branch; precision matters more than speed.
- Data scientists: practise presenting a model's output to a non-technical audience after a 15-minute preparation window.
- Have stories ready for challenging the status quo and delivering under pressure; the behavioral round is highly structured.
- Ask early whether you are hiring into a team or a pool, and budget time for team matching after the Power Day.
See the rest of the Capital One loop
The later stages, the questions reported for them and how to prepare are in Pro.
Unlock with ProPractice the Capital One loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Capital One interview FAQ
How many interview rounds does Capital One have for ML roles?
Candidates report 9 stages: Recruiter screen, Online coding assessment, Data science assessment (data scientist track), Hiring manager screen, Power Day: ML technical & job fit, Power Day: ML system design, Power Day: case interview, Power Day: statistics role-play (data scientist track) and Power Day: behavioral & job fit. A recruiter call, then a 70-minute online coding assessment of four questions for ML and AI engineers (data scientists get a 90-minute pandas and modelling assessment or a week-long take-home), and for many candidates a 30-60 minute hiring-manager screen, before or after the assessment. Several recent ML engineer reports went straight from the assessment to the final stage. The final stage is a Power Day of four one-hour video interviews on one day: for ML engineers usually ML technical, ML system design, a case and behavioral; for data scientists technical, case, a statistics role-play and job fit. Rounds are scored separately and combined. Team matching can follow the Power Day and add weeks.
How long does the Capital One ML interview process take?
Reported timelines run 1-3 months.
What does Capital One ask in ML interviews?
By reported question mix: General coding 28%; ML concepts & fundamentals 22%; Business case 15%; Behavioral & job fit 12%; ML system design & deployment 10%; Statistics & model interpretation 8%. The most repeated themes: Four-question timed assessment: two warm-ups, a tedious simulation, a harder counting problem; Resume and project deep dive with the hiring manager; ML fundamentals rapid-fire: bias-variance, overfitting, boosting vs bagging.
How should I prepare for the Capital One ML interview?
Treat the coding assessment as the main filter for ML and AI engineers: practise four mixed problems in 70 minutes with no autocomplete and no copy and paste. Drill implementation-heavy problems (grid paths, matrix rotation, placement simulations, interval sweeps) more than advanced algorithms. Prepare crisp, example-backed answers to the classic ML fundamentals; the technical rounds cover many topics quickly.
Are there free Capital One interview questions?
Yes. LeetML has 31 questions reported from Capital One interviews, and 3 of them are free to solve in the browser; they are listed by stage above.
Related interview guides
- JPMorgan ChaseA recorded first-round assessment, then team interviews that test ML basics, your own projects and, for quant tracks, probability and statistics.
- American ExpressTree-based ML fundamentals, SQL, business cases and guesstimates, with heavy weight on your own projects.
- PayPalSQL and Python screens, medium-difficulty algorithm coding, ML fundamentals and fraud-flavoured cases, with loops that vary by team.
- StripePractical multi-part coding, a timed dataset-to-model round, an ML bug squash and payments-flavoured ML design.
- VisaSQL and Python screens, ML fundamentals, and consulting-style cases built on payments data.