Company interview guide
Cursor Machine Learning Interview Guide (2026)
Short coding screens inside a real repository, then a project day where you scope, build and defend a feature in Cursor's own code.
Cursor's machine learning interview loop has 5 reported stages (Hiring-manager or recruiter call, Technical interviews: coding in a repository, Project deep dive and hiring-manager round, Project onsite, Possible design and ML depth questions).
Roles covered: ML Engineer, AI Engineer, LLM Engineer and Research Scientist. The question bank holds 12 questions reported from Cursor interviews.
- Stages
- 5
- Reports linked to bank questions
- 12
- Bank report dates
- Sep 2026–Sep 2026
- Questions in bank
- 12 (0 free)
- Last reviewed
- October 2026
How the Cursor interview loop works
A conversation with a hiring manager or recruiter, then two to four technical interviews of about an hour (coding inside a repository, and in one senior ML report also a project deep dive and a hiring-manager style round), then a project onsite. The onsite is described by the company as in-office; one new-grad candidate did a remote version of about eight to nine hours with a kickoff, a Slack channel, a mid-day check-in and a 30-minute presentation. One candidate reports reference and backchannel checks after the onsite. The outline rests on three first-hand reports plus the company's postings, so treat it as typical rather than fixed.
The Cursor interview process, stage by stage
01Hiring-manager or recruiter call
About 45 min conversation; some candidates speak with a hiring manager before or instead of a recruiter
An informal call about your background, past projects and why you want to work on Cursor. One new-grad candidate had this call with the hiring manager directly, who was frank about a flat, fast-moving and intense company; a frontend candidate spoke with the hiring manager first and the recruiter second.
What it tests
- Why Cursor, and how you use the product day to day
- How your past work maps to the team (editor, agents, model training, inference, infrastructure)
- Whether you want a small, intense, in-person team
- Logistics: level, location (San Francisco or New York), timeline
How to prepare
- Use Cursor heavily for a week beforehand and bring specific opinions on what works and what does not.
- Ask which technical rounds you will get, whether AI tools are allowed in each and whether the onsite is in person or remote.
- Have a short story ready about one project you are proud of; the application form asks for it too.
Two candidate reports (a new-grad engineer and a frontend engineer) describe an opening hiring-manager conversation; neither is an ML role, but this call is shared by every track.
02Technical interviews: coding in a repository
Two to three short technicals per the company; about 60 min each, often in a cloned repository with screen sharing
Practical coding framed around something the product needs, written in a real or realistic codebase. Both reports that name the problem built a Merkle or hash tree over repository files and used it to find what changed; a senior ML candidate also had to speed up an existing Python library by swapping in better data structures. AI-tool rules differ: the CEO said in 2025 that the first screen allows nothing beyond autocomplete, while one candidate could look up syntax with AI tools if he said what he was searching for.
What it tests
- Hash trees over files and directories, and diffing two snapshots without comparing every file
- Reading unfamiliar code and file layouts before you write
- Choosing data structures to speed up existing code, and explaining the complexity gain
- Engineering judgment over perfect syntax
How to prepare
- Practise building a Merkle tree from a path-to-content map and returning changed files top-down.
- Rehearse optimizing someone else's slow Python: profile, find the hot loop, change the structure, measure.
- Practise without AI help in case your round bans it, and narrate any lookup you make when it is allowed.
- Set up a clean local environment beforehand so cloning and running the repo costs no time.
Cursor's postings, ML Platform and Model Routing included, list 2-3 short technicals (one senior ML candidate had four); two candidate reports (senior MLE in 2026, new-grad engineer) describe repository coding with a hash-tree task.
03Project deep dive and hiring-manager round
About 60 min each; in one senior ML loop these were two of the four screening interviews
A detailed walk through a past project, plus a broad conversation similar to a hiring-manager screen. Only one ML report describes these as separate rounds, and that candidate did not reach the onsite, so it is unclear how often they appear.
What it tests
- Depth on one past project: goals, design decisions, results, what you would change
- Your own contribution versus the team's
- Ownership and working with little direction
- Fit with a small team that expects self-motivated work
How to prepare
- Pick one ML or systems project and prepare its numbers: scale, latency or quality gains, and cost.
- Expect follow-ups on trade-offs; be ready to say what failed and why.
- Link the project to a part of Cursor you would work on, such as retrieval, completions or agents.
A single senior MLE report from 2026 lists a project deep dive and a hiring-manager style round among four screens; no other account confirms them as separate rounds.
04Project onsite
At the office per the company and the CEO (up to two days with the team); one candidate did a remote day of 8-9 hours ending in a 30-min presentation
The deciding round. You get a slice of the Cursor codebase, pick something worth building, scope it, build it largely on your own and present it. In the remote version one candidate had a kickoff, a Slack channel for questions and a mid-day check-in, and the closing questions were about why he chose that feature, whether it already existed in the code and the trade-offs of what he added. The CEO describes the in-person version as working with the team, sharing meals and demoing at the end. Whether the day is paid is not confirmed by any first-hand report.
What it tests
- Finding your way around a large unfamiliar codebase quickly
- Choosing a feature that matters and cutting it to fit the time
- Building with little direction, using docs, READMEs and AI tools sensibly
- Explaining and defending every decision and every line you commit
- Genuine interest in the product and the problem space
How to prepare
- Spend the first hour reading and tracing a request end to end before choosing what to build.
- Ship a small working core early, then extend it; a finished slice beats an ambitious half-build.
- Use AI tools freely if allowed, but read and test what they produce; reviewers probe whether you understand it.
- Ask your recruiter whether the day is paid, in person or remote, and what the IP terms are.
Cursor's postings, ML Platform and Model Routing included, describe an in-office onsite with a small project; one new-grad report details a remote 8-9 hour version and a frontend candidate calls it very hard. No ML account yet.
05Possible design and ML depth questions
Not confirmed as a separate round; questions on editor, retrieval and serving design may come up in technicals, the deep dive or the project review
Public question lists for ML and AI roles at Cursor center on the product itself: instant next-edit suggestions, indexing a large monorepo, applying model edits reliably, safe multi-file agents, evaluating code-editing models and cutting serving cost. No first-hand account we hold describes a standalone design round, so prepare these as depth for any conversation rather than as a fixed interview.
What it tests
- Latency budgets for an interactive completion or next-edit model
- Code retrieval: chunking, embedding, keeping an index fresh as files change
- Turning model output into safe, fast edits and multi-file agent changes
- Offline and online evaluation of code-editing models
- Inference cost and the main levers to reduce it
How to prepare
- Learn how Cursor describes its own features (tab, indexing, apply, agents) in its blog and docs.
- For each design, give the latency or cost target first, then the architecture, then how you would measure it.
- Connect retrieval answers to the Merkle-tree idea: hash-based change detection keeps re-indexing cheap.
No first-hand report describes a design round; the questions come from a compiled public write-up, and one guide says some loops add design or ML conversations.
Questions and themes reported most often
- Build a Merkle or hash tree over repository files and find what changed (2 reports, technical interviews: coding in a repository)
- Opening conversation with the hiring manager about background and why Cursor (2 reports, hiring-manager or recruiter call)
- Scope, build and present a feature in a slice of the real codebase (1 reports, project onsite)
- Speed up an existing Python library with better data structures (1 reports, technical interviews: coding in a repository)
- Deep dive on a past project (1 reports, project deep dive and hiring-manager round)
What Cursor asks about
- Practical coding in a real codebase35%
- Project day: scope, build, present30%
- Editor and LLM system design20%
- Project deep dive & motivation15%
Cursor interview prep checklist
- Built a Merkle tree over a file snapshot and returned changed files without comparing every file
- Optimized a slow Python module by changing its data structures, and measured the gain
- Shipped and presented a scoped feature in an unfamiliar open-source codebase in one day
- Can explain each decision and every line of code in that build
- Can sketch a low-latency next-edit or code-retrieval design with a metric for each part
- Have a crisp, personal answer to why Cursor, backed by real use of the product
Tips
- Make the Merkle-tree diff your first coding drill; it is the only problem more than one candidate reports.
- Practise coding inside a repository you did not write: clone, run, read, then change.
- Prepare for both AI rules: a screen with nothing beyond autocomplete, and rounds where AI help is allowed if you narrate it.
- Rehearse a full-day build on an open-source codebase: pick a feature, scope it, ship it, present it in 30 minutes.
- Be ready to justify every choice in your project, including why the feature matters and what already existed.
- Use Cursor daily before you apply and form concrete views on its features and rivals.
- For ML and research roles, prepare depth on retrieval for code, low-latency serving and evaluating code models.
- Expect a small, intense, in-person team and reference checks after the onsite.
Practice the Cursor loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Cursor interview FAQ
How many interview rounds does Cursor have for ML roles?
Candidates report 5 stages: Hiring-manager or recruiter call, Technical interviews: coding in a repository, Project deep dive and hiring-manager round, Project onsite and Possible design and ML depth questions. A conversation with a hiring manager or recruiter, then two to four technical interviews of about an hour (coding inside a repository, and in one senior ML report also a project deep dive and a hiring-manager style round), then a project onsite. The onsite is described by the company as in-office; one new-grad candidate did a remote version of about eight to nine hours with a kickoff, a Slack channel, a mid-day check-in and a 30-minute presentation. One candidate reports reference and backchannel checks after the onsite. The outline rests on three first-hand reports plus the company's postings, so treat it as typical rather than fixed.
What does Cursor ask in ML interviews?
By reported question mix: Practical coding in a real codebase 35%; Project day: scope, build, present 30%; Editor and LLM system design 20%; Project deep dive & motivation 15%. The most repeated themes: Build a Merkle or hash tree over repository files and find what changed; Opening conversation with the hiring manager about background and why Cursor; Scope, build and present a feature in a slice of the real codebase.
How should I prepare for the Cursor ML interview?
Make the Merkle-tree diff your first coding drill; it is the only problem more than one candidate reports. Practise coding inside a repository you did not write: clone, run, read, then change. Prepare for both AI rules: a screen with nothing beyond autocomplete, and rounds where AI help is allowed if you narrate it.
Are there free Cursor interview questions?
LeetML has 12 questions reported from Cursor interviews. The stage sections above link the free practice problems for the topics each round covers; the reported-question set itself is in the app.
Related interview guides
- CognitionHard take-home coding first, then practical rounds built around coding agents; very few first-hand reports.
- OpenAIPractical coding, hands-on ML debugging and LLM system design, with a strong mission check.
- AnthropicA timed, level-by-level build exercise, practical live coding, LLM infrastructure design and a values interview that really counts.
- PerplexityPractical multi-part coding, retrieval and LLM depth, a hiring-manager project deep dive and a real check on product taste.
- DatabricksPractical coding, LLM-centred ML design on data-platform problems, and a long, structured panel.
Official sources
- Cursor careers (official)
- Software Engineer, Product posting with the interview outline (official)
- Cursor: Software Engineer, ML Platform posting with the interview outline (official)
- Cursor: Software Engineer, Model Routing & Inference posting with the interview outline (official)
- Cursor: Software Engineer, New Grad 2027 posting with the interview outline (official)