Company interview guide

Applied Intuition Machine Learning Interview Guide (2026)

Coding-heavy screens, hands-on transformer debugging and a fast loop, for an autonomy software company that is open about its pace.

Applied Intuition's machine learning interview loop has 5 reported stages (Recruiter or hiring-manager call, Technical screen: coding, Technical screen: ML code debugging, Final loop: coding and AI / ML system design, Behavioral and manager chat); typical length: 2-6 weeks.

Roles covered: ML Engineer and Research Scientist. The question bank holds 7 questions reported from Applied Intuition interviews, 3 of them free to solve in your browser.

Try a free question

Interview process length
2-6 weeks
Stages
5
Reports linked to bank questions
7
Bank report dates
Sep 2025–Apr 2026
Questions in bank
7 (3 free)
Last reviewed
October 2026

How the Applied Intuition interview loop works

A recruiter or hiring-manager call, then one technical screen of about 45-60 minutes, then a final loop. ML candidates describe the screen as either a coding problem (an algorithm question, a deque, a config validator) or debugging a small transformer or vision model in a notebook until it trains. The one ML report that reached the end described four more rounds: two coding, one AI / ML system design and one behavioral. Software engineer reports from 2025-2026 add that new-grad candidates may get an online assessment first, and that onsites can include several coding rounds, an AI-assisted coding segment, system design and a manager chat. Two reports say recruiters were upfront about roughly 50-hour weeks. Most stages here rest on fewer than five ML-role reports.

The Applied Intuition interview process, stage by stage

01Recruiter or hiring-manager call

About 30 min with a recruiter, or a first conversation with the hiring manager

Some candidates speak with a recruiter about background, languages and ML frameworks; others start with the hiring manager discussing past work and research interests. Recruiters have been direct about the intensity of the job, including roughly 50-hour weeks and hard coding rounds, so use the call to decide whether that suits you.

What it tests

  • A clear summary of your past ML or research work and its impact
  • Which languages and ML frameworks you use day to day
  • How your interests fit the team's roadmap
  • Comfort with the pace and in-office expectations

How to prepare

  • Prepare a two-minute walkthrough of your most relevant project, including a measurable result.
  • Ask whether your technical screen will be algorithm coding or ML code debugging; reports show both.
  • Ask about expected hours and location early; recruiters have raised them on their own.

Recruiter or hiring-manager first call in 6 of 8 ML-role reports (2023-2026) plus 4 software engineer reports from 2025-2026; two mention a 50-hour week.

02Technical screen: coding

About 45-60 min of live coding with screen share; one medium to hard problem, sometimes in C++ for engineering roles

The most common first technical round for ML roles is a general coding problem. Reported prompts range from a hard binary-search variant to practical tasks with custom classes, such as validating nested configs against hand-written descriptors or building a fixed-capacity deque. Interviewers often give hints, but candidates who missed edge cases were rejected weeks later.

What it tests

  • Solving a medium to hard algorithm problem cleanly within the slot
  • Implementing a small practical component with your own data structures
  • Agreeing assumptions and edge cases with the interviewer before coding
  • Explaining your approach while you code

How to prepare

  • List edge cases out loud (missing fields, wrong types, nested lists, full or empty buffers) and handle them explicitly.
  • Practise grid BFS variants and binary search on the answer; both recur in engineering screens.
  • Expect to write the code from an empty editor with no tests provided, so test it yourself before saying you are done.

General coding screen in 5 of 8 ML-role reports from 2024-2026 (one adds a GPU question); software engineer reports describe the same 45-minute format but are not counted here.

03Technical screen: ML code debugging

About 60 min in a notebook with PyTorch model, training and inference code; some teams ask a GPU systems question instead

Some ML and research teams replace the algorithm problem with a notebook containing a small model that does not train. One candidate had to find three bugs in a minimal GPT (attention mask, training step, missing positional encoding) to push loss below a target, with a KV cache follow-up expected; another debugged vision transformer training code. A 2026 report paired coding with a question on making a GPU reduction deterministic.

What it tests

  • Spotting a wrong causal or padding mask
  • Finding errors in the training step (loss, label shift, optimizer calls)
  • Writing positional encodings from memory
  • Adding a KV cache that matches uncached output
  • Reasoning about floating-point order in parallel reductions

How to prepare

  • Be able to write sinusoidal and learned positional encodings without looking them up.
  • Check mask direction, the one-token label shift and missing zero_grad first; they are the usual planted bugs.
  • Know why float addition is not associative and how a fixed tree or scan order makes a reduction reproducible.

Free reported questions for this stage

Notebook debugging of model training code in 2 ML-role reports from 2025 (a GPT and a vision transformer), plus one 2026 report with a GPU reduction question.

04Final loop: coding and AI / ML system design

About four back-to-back interviews; in the one ML report, two coding rounds and one AI / ML system design round

Only one ML-role report describes the final loop: two more coding rounds and one AI / ML system design round, alongside a behavioral interview. Software engineer onsites in 2025-2026 are similar but can run three to five coding rounds, a system design round and an AI-assisted coding segment, with precise handling of corner cases expected. Ask your recruiter for the exact mix.

What it tests

  • Medium to hard coding with careful corner-case handling
  • Designing an ML system end to end, including data, training, evaluation and serving
  • Latency and reliability trade-offs for onboard or simulation workloads
  • Working effectively with AI coding tools when a round allows them

How to prepare

  • Practise two or three coding problems back to back to rehearse the pace of a full day.
  • Prepare one ML system design for a perception or driving-data pipeline, since that is the company's domain.
  • If a round allows AI assistance, still explain and verify every line it produces.

One 2024 ML-role report lists two coding rounds and an AI / ML design round; several 2025-2026 software engineer reports describe similar onsites but are not counted.

05Behavioral and manager chat

Usually one conversation of up to an hour with a manager or leader, often at the end of the loop

Expect a deep dive on your most impactful project and a direct question on why you want this role. A few candidates found the manager round probing about career choices, so have a calm, honest account of your path.

What it tests

  • Depth and ownership in your most impactful project
  • Why Applied Intuition and why this team
  • How you have made past career decisions
  • Working under pressure in a fast-moving team

How to prepare

  • Pick one project and rehearse follow-ups on your own contribution, trade-offs and results.
  • Tie your answer on motivation to autonomy, simulation or the specific product the team builds.

One ML-role report from 2024 plus 4 software engineer reports from 2025-2026 describe a behavioral or manager round; this stage is shared across tracks.

Questions and themes reported most often

  • Walk through your research or most impactful project (3 reports, recruiter or hiring-manager call)
  • Grid BFS variants, including multi-source spread and choosing the best cell (software engineer reports) (3 reports, technical screen: coding)
  • Debug a small transformer's training code until loss drops, then add a KV cache (2 reports, technical screen: ml code debugging)
  • Why do you want to work here (2 reports, behavioral and manager chat)
  • Validate a nested config against custom descriptor classes (1 reports, technical screen: coding)
  • Fixed-capacity double-ended queue with O(1) operations (1 reports, technical screen: coding)
  • Hard binary-search variant (1 reports, technical screen: coding)
  • Deterministic parallel reduction on a GPU, prefix-sum style (1 reports, technical screen: ml code debugging)

What Applied Intuition asks about

  • General coding40%
  • ML coding & debugging30%
  • Behavioral & past work15%
  • AI / ML system design10%
  • GPU & performance5%

Applied Intuition interview prep checklist

  • Fixed a buggy minimal GPT so it trains below a loss target, without notes
  • Wrote positional encodings and a causal plus padding mask from scratch
  • Added a KV cache and checked it against uncached generation
  • Solved a practical implementation task with nested types and stated every assumption
  • Solved a grid BFS problem and a binary-search variant inside 45 minutes
  • Prepared a deep dive on one project and a specific answer to why Applied Intuition
  • Asked the recruiter for the format of every round and the expected hours

Tips

  • Ask the recruiter which screen you will get: algorithm coding or ML code debugging. Prepare for both if unsure.
  • Drill a minimal GPT until you can fix its mask, training step and positional encoding quickly, then add a KV cache.
  • Practise practical implementation problems with your own classes, where clarifying assumptions matters as much as the algorithm.
  • Get comfortable with grid BFS and binary-search variants under a 45-minute limit.
  • Handle edge cases explicitly; several rejected candidates said they solved the core problem but missed cases.
  • Know the basics of GPU numerics: why float sums differ by order and how to make reductions reproducible.
  • Prepare one ML system design grounded in driving or sensor data, with evaluation and latency in mind.
  • Go in with clear eyes about the pace: recruiters describe long weeks and demanding coding rounds.

See the rest of the Applied Intuition loop

The later stages, the questions reported for them and how to prepare are in Pro.

Unlock with Pro

Practice the Applied Intuition loop

Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.

Applied Intuition interview FAQ

How many interview rounds does Applied Intuition have for ML roles?

Candidates report 5 stages: Recruiter or hiring-manager call, Technical screen: coding, Technical screen: ML code debugging, Final loop: coding and AI / ML system design and Behavioral and manager chat. A recruiter or hiring-manager call, then one technical screen of about 45-60 minutes, then a final loop. ML candidates describe the screen as either a coding problem (an algorithm question, a deque, a config validator) or debugging a small transformer or vision model in a notebook until it trains. The one ML report that reached the end described four more rounds: two coding, one AI / ML system design and one behavioral. Software engineer reports from 2025-2026 add that new-grad candidates may get an online assessment first, and that onsites can include several coding rounds, an AI-assisted coding segment, system design and a manager chat. Two reports say recruiters were upfront about roughly 50-hour weeks. Most stages here rest on fewer than five ML-role reports.

How long does the Applied Intuition ML interview process take?

Reported timelines run 2-6 weeks.

What does Applied Intuition ask in ML interviews?

By reported question mix: General coding 40%; ML coding & debugging 30%; Behavioral & past work 15%; AI / ML system design 10%; GPU & performance 5%. The most repeated themes: Walk through your research or most impactful project; Grid BFS variants, including multi-source spread and choosing the best cell (software engineer reports); Debug a small transformer's training code until loss drops, then add a KV cache.

How should I prepare for the Applied Intuition ML interview?

Ask the recruiter which screen you will get: algorithm coding or ML code debugging. Prepare for both if unsure. Drill a minimal GPT until you can fix its mask, training step and positional encoding quickly, then add a KV cache. Practise practical implementation problems with your own classes, where clarifying assumptions matters as much as the algorithm.

Are there free Applied Intuition interview questions?

Yes. LeetML has 7 questions reported from Applied Intuition interviews, and 3 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources