Company interview guide

Perplexity Machine Learning Interview Guide (2026)

Practical multi-part coding, retrieval and LLM depth, a hiring-manager project deep dive and a real check on product taste.

Perplexity's machine learning interview loop has 8 reported stages, from recruiter and intro calls to founder or leader conversation and decision; typical length: 1-4 weeks.

Roles covered: ML Engineer, AI Engineer, LLM Engineer and Research Scientist. The question bank holds 22 questions reported from Perplexity interviews.

Try a free question

Interview process length
1-4 weeks
Stages
8
Reports linked to bank questions
24
Bank report dates
Apr 2025–Sep 2026
Questions in bank
22 (0 free)
Last reviewed
October 2026

How the Perplexity interview loop works

Recruiter call (sometimes a short call with a hiring manager or senior leader that can turn technical), then a technical screen of about an hour, then an onsite of four to five interviews: practical coding, ML and search depth, a design or product conversation and a hiring-manager deep dive with values questions. A founder or leader conversation follows, and the company says it decides within a week of the onsite. ML-role reports are few and the round mix varies by team, so treat this as a typical outline rather than a fixed loop.

The Perplexity interview process, stage by stage

01Recruiter and intro calls

Recruiter or head-recruiter call; some candidates also get a short call with a hiring manager or senior leader

The recruiter covers your experience, what you want next and why Perplexity. Some candidates then have a brief hiring-manager call on motivation; one research engineer's 15-minute 'casual' call with a senior leader was fully technical, and one ML candidate was asked about positional encoding in an early screen. A strong candidate with other onsites under way was moved straight to the onsite.

What it tests

  • Why Perplexity, and how you use the product
  • How your past work fits search, retrieval, ranking or LLM work
  • For research roles, a quick technical probe such as how to post-train a base model to follow instructions
  • Basic transformer knowledge, even in a call billed as non-technical

How to prepare

  • Treat any call with an engineer or leader as a technical interview, even if it is described as a chat.
  • Use the product heavily beforehand and have two or three specific opinions about what works and what does not.
  • Ask which onsite rounds apply to your role; the mix of backend and ML rounds varies by team.

Five ML or research reports since 2024 describe a recruiter or leader intro call, two of them technical; three engineering reports describe the same call. The company's interview guide lists a recruiter phone screen.

02Technical screen

About 60 min live coding, often multi-part; some engineering candidates get an online assessment instead

The company describes a standard programming interview for engineers. ML-role reports mention a practical coding task on deduplication and optimization, and another with Python and scikit-learn plus ML fundamentals and how to improve an evaluation metric. Engineering candidates describe hour-long screens with three or four escalating parts, such as an in-memory file system, and a few had a hard timed online assessment.

What it tests

  • Working, clean Python under time pressure, part by part
  • Efficient algorithms for data tasks such as near-duplicate detection
  • Training and evaluating a simple model with standard libraries and explaining how to improve it
  • Adapting quickly when the interviewer adds a requirement

How to prepare

  • Practise multi-part problems where each part extends the last, and keep earlier parts passing.
  • Be fluent with hashing and shingling ideas (MinHash, locality-sensitive hashing) for dedup questions.
  • For ML screens, be ready to fit a scikit-learn model quickly, pick a metric and propose concrete improvements.

Two ML-role phone-screen reports (dedup coding; scikit-learn plus ML basics). Engineering reports describe similar multi-part screens; not counted. The company guide lists a technical screen, usually a programming interview.

03Onsite: practical coding

One or two 45-60 min rounds of live coding, sometimes with pseudocode or a debugging exercise

Expect to build or fix a small, stateful piece of software while the interviewer adds parts, not a classic puzzle. One ML hire had three 45-minute interviews mixing backend and ML with pseudocode, live coding and debugging; another MTS candidate had two coding rounds around a core-expertise round. Engineering reports mention tokenizers, key-value stores with timestamps and in-memory file systems; public lists add async API batching and beam search.

What it tests

  • Clean, state-aware code that survives new requirements
  • Debugging unfamiliar code quickly
  • Concurrency limits, retries and per-item error handling in API pipelines
  • Decoding algorithms such as beam search
  • Explaining trade-offs out loud as you go

How to prepare

  • Drill plain-Python, standard-library implementations of caches, stores and tokenizers with several follow-up parts.
  • Get comfortable with asyncio semaphores, backoff and isolating failures.
  • If you fall behind, say what the next part needs; candidates who finished only half the parts report no offer.

Two MTS reports (2025, 2026) describe onsite coding with pseudocode or debugging; engineering reports agree on multi-part, stateful coding but are not counted. The company guide lists 4-5 onsite interviews but no coding round.

04Onsite: ML, search and LLM depth

45-60 min technical conversation on your core area, sometimes with pseudocode

A round on the building blocks of search and LLM products, or on your own specialty. One MTS candidate called it a core-expertise round and needed hints; another had ML topics mixed with backend ones. Public question lists add reranker choices, choosing passages for a limited context window, citation verification, decoding and latency debugging; those are not first-hand.

What it tests

  • Bi-encoders, cross-encoders, late interaction and LLM rerankers, with their costs
  • Selecting and ordering retrieved passages under a token budget
  • Claim-level attribution and citation checks
  • Serving metrics such as time to first token, and how to debug a regression
  • Post-training basics: instruction tuning and preference optimization

How to prepare

  • Be ready to put numbers on things: candidates per stage, latency per stage, tokens in context.
  • For a latency regression, separate queueing, prefill and decode time first.
  • Go deep on whatever the job post lists; the company says any skill on your application is fair game.

Two MTS reports describe an ML or core-expertise round; specific search and LLM topics come from a compiled public question list and prep guides, not first-hand accounts. The company guide does not name this round.

05Onsite: retrieval and answer-engine design

About 45-60 min open discussion; some variants ask you to improve an existing system

Design questions are about Perplexity's own product. The one first-hand ML question was search personalization when user interests drift: how to scope it, how fast to forget old signals and how to evaluate it. Public lists add a cited, streamed answer engine, web-scale retrieval with freshness, breaking news within minutes and routing across LLM providers.

What it tests

  • Scoping an open product problem and choosing evaluation up front
  • Hybrid sparse and dense retrieval with reranking
  • Index freshness and handling recent events
  • End-to-end latency budgets across retrieval, reranking and generation
  • Routing, caching and failover across model providers

How to prepare

  • Have a ready blueprint for a cited answer engine and split its latency budget per stage.
  • Name the metric for every component, from retrieval recall to citation accuracy.
  • For personalization, discuss decay of old signals, exploration and offline versus online evaluation.

One ML-role report gives a design question (search personalization). Engineering onsite reports often include system design; answer-engine prompts come from public lists. The company guide does not name a design round.

06Onsite: product and answer quality

45-60 min, sometimes with a product manager or growth leader

Product taste is checked. One hire met a product manager about challenging collaboration and the head of growth, who asked how to measure the quality of the product and how to size the impact of a Pro partnership. The company guide asks candidates to use the product and spot where they could improve its quality.

What it tests

  • Offline and online evaluation of answer quality with automated and human signals
  • Designing and calibrating an LLM judge
  • Sizing the business impact of a product decision
  • Opinions on where Perplexity beats or loses to traditional search

How to prepare

  • Prepare a layered evaluation plan: golden sets, LLM judges checked against human labels, then online metrics.
  • Bring concrete examples of good and bad answers you have seen in the product.

One MTS onsite report (2024) with product-manager and growth-leader rounds; public guides also describe a product round. The company guide asks for product use but does not list a round.

07Hiring-manager deep dive and values

Deep dive or presentation on a past project with the hiring manager (one engineer reported 90 min), plus a values and communication interview

The company guide says the onsite includes a detailed deep dive with the hiring manager on your past work, and that team matching happens here. One hire presented a technical project from a previous job, then had a values and communication interview; behavioral questions included disagreeing with company strategy.

What it tests

  • Your personal contribution and depth behind one project
  • Ownership, speed and working through ambiguity
  • Collaboration and handling disagreement
  • Fit with a specific team

How to prepare

  • Pick one project and prepare its motivation, key decisions, measured impact and what you would change.
  • Be precise about what you did yourself; the company asks for personal contributions, not team results.
  • Have a story about disagreeing with a direction and what you did about it.

Two MTS reports (presentation and values rounds; a hiring-manager call) and two engineering reports of an HM deep dive. The company guide puts a deep dive on past work in the onsite, with team matching; one report had it early.

08Founder or leader conversation and decision

Conversation with a founder or senior leader after the onsite; decision within about a week

The company guide lists a final interview with a founder or leader and a decision within a week of the onsite. One hire heard positive feedback within 24 hours and then had a call with the CEO about the business and roadmap, closer to a sell call than an interview.

What it tests

  • Conviction about the product and the mission
  • Judgment and clarity when discussing strategy
  • Questions that show you have thought about the business

How to prepare

  • Prepare questions about the roadmap and how your team would move key product metrics.
  • Expect speed: be ready to discuss timelines and competing offers openly.

One MTS report of a CEO call after positive feedback within a day; the company guide lists a final interview with a founder or leader and a decision within a week of the onsite.

Questions and themes reported most often

  • Multi-part stateful coding, such as an in-memory file system or a timestamped key-value store (5 reports, technical screen)
  • Presentation or long deep dive on a past technical project (2 reports, hiring-manager deep dive and values)
  • Mixed backend and ML interviews with pseudocode, live coding and debugging (2 reports, onsite: practical coding)
  • Byte-level or BPE tokenizer implemented from scratch (2 reports, onsite: practical coding)
  • Why Perplexity and why this role (1 reports, recruiter and intro calls)
  • Post-train a base LLM so it follows instructions (1 reports, recruiter and intro calls)
  • Positional encoding in transformers, asked in an early screen (1 reports, recruiter and intro calls)
  • Detect near-duplicates efficiently in a coding screen (1 reports, technical screen)

What Perplexity asks about

  • Practical coding26%
  • Retrieval, ranking & LLM concepts24%
  • ML / LLM system design22%
  • Behavioral & project deep dive16%
  • Evaluation & product sense12%

Perplexity interview prep checklist

  • Solved three multi-part stateful coding problems (store, file system, tokenizer) in plain Python under time
  • Implemented near-duplicate detection and beam search from scratch
  • Can design a cited answer engine with a per-stage latency budget and a freshness plan
  • Can compare cross-encoder, late-interaction and LLM rerankers on cost and quality
  • Have a layered plan for evaluating answer quality offline and online
  • Prepared a deep dive on one past project, including what did not work
  • Have specific, honest opinions on the product and why you want to work there

Tips

  • Expect practical, multi-part coding: build something small, then extend it as the interviewer adds parts.
  • Use the product daily for a week before the loop and note specific failures and fixes; the company asks you to.
  • Treat every call with an engineer or leader as technical, including ones described as casual.
  • Prepare one past project in depth, with your own contribution separated from the team's.
  • Know retrieval, reranking, citation and serving-latency trade-offs well enough to put numbers on them.
  • Revise anything listed on your application; the company says it is fair game.
  • For research roles, be ready to discuss post-training (instruction tuning, preference optimization) from first principles.
  • Expect a fast process and be ready to move quickly on decisions.

Practice the Perplexity loop

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

Perplexity interview FAQ

How many interview rounds does Perplexity have for ML roles?

Candidates report 8 stages: Recruiter and intro calls, Technical screen, Onsite: practical coding, Onsite: ML, search and LLM depth, Onsite: retrieval and answer-engine design, Onsite: product and answer quality, Hiring-manager deep dive and values and Founder or leader conversation and decision. Recruiter call (sometimes a short call with a hiring manager or senior leader that can turn technical), then a technical screen of about an hour, then an onsite of four to five interviews: practical coding, ML and search depth, a design or product conversation and a hiring-manager deep dive with values questions. A founder or leader conversation follows, and the company says it decides within a week of the onsite. ML-role reports are few and the round mix varies by team, so treat this as a typical outline rather than a fixed loop.

How long does the Perplexity ML interview process take?

Reported timelines run 1-4 weeks.

What does Perplexity ask in ML interviews?

By reported question mix: Practical coding 26%; Retrieval, ranking & LLM concepts 24%; ML / LLM system design 22%; Behavioral & project deep dive 16%; Evaluation & product sense 12%. The most repeated themes: Multi-part stateful coding, such as an in-memory file system or a timestamped key-value store; Presentation or long deep dive on a past technical project; Mixed backend and ML interviews with pseudocode, live coding and debugging.

How should I prepare for the Perplexity ML interview?

Expect practical, multi-part coding: build something small, then extend it as the interviewer adds parts. Use the product daily for a week before the loop and note specific failures and fixes; the company asks you to. Treat every call with an engineer or leader as technical, including ones described as casual.

Are there free Perplexity interview questions?

LeetML has 22 questions reported from Perplexity 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

Official sources