Company interview guide
Harvey Machine Learning Interview Guide (2026)
Legal AI company whose ML loop appears to centre on grounded retrieval, evaluation and a project deep dive; evidence is thin.
Harvey's machine learning interview loop has 6 reported stages (Recruiter screen or direct outreach, Paired coding, Architecture and ML / LLM design, Evaluation and grounding, Customer or partner scenario, Behavioral and project deep dive); typical length: Not reported.
Roles covered: ML Engineer, AI Engineer, LLM Engineer and Forward Deployed Engineer. The question bank holds 10 questions reported from Harvey interviews.
- Interview process length
- Not reported
- Stages
- 6
- Reports linked to bank questions
- 10
- Bank report dates
- Jun 2026–Sep 2026
- Questions in bank
- 10 (0 free)
- Last reviewed
- October 2026
How the Harvey interview loop works
First contact may come from a recruiter or, in one 2024 report, directly from a founder who skipped the phone screen entirely. That candidate's onsite had four interviews, about half technical and the rest behavioral. A compiled 2026 write-up describes a longer sequence for software, ML, forward-deployed and research engineers: recruiter screen, a paired coding round, a presentation of an architecture for a legal workflow, a round on designing evaluations, and a customer or partner scenario. A separate 2026 ML engineer onsite included an open discussion of reinforcement learning for long-horizon agents. The company's own pages do not describe the interview process. This outline rests on two first-hand reports plus one compiled write-up, so expect the rounds to differ by team and level.
The Harvey interview process, stage by stage
01Recruiter screen or direct outreach
A recruiter call in the compiled outline; one candidate was invited by a founder straight to the onsite with no screen
How the process starts seems to vary. A compiled write-up lists a recruiter screen first, while the one detailed candidate account had no call at all: a founder emailed and invited them to an onsite. Use whatever first contact you get to ask which rounds follow.
What it tests
- Why Harvey and why legal AI
- How your past work maps to the team you are targeting
- Logistics: level, location, timeline
How to prepare
- Ask for the list of rounds and their format; public information on the loop is sparse.
- Legal experience is not required, but show curiosity about how lawyers actually work.
- If you are invited straight to an onsite, ask what the technical sessions cover before you go.
One ML engineer report from 2024 describes founder outreach straight to an onsite with no screen; a compiled public write-up lists a recruiter screen instead.
02Paired coding
Pair programming with an interviewer; length and tooling not reported
The compiled outline lists a paired coding round with a document-processing flavour, for example writing a chunker that never cuts a clause and keeps each chunk understandable on its own. No first-hand report yet describes this round, so treat the example as indicative.
What it tests
- Clean, readable code written while talking through choices
- Text processing with boundary rules and overlap
- Handling edge cases such as very long units or empty input
How to prepare
- Practise chunking and parsing tasks over structured documents, not only classic algorithms.
- State your invariants (never split a unit, fixed overlap) before you code.
No first-hand report describes this round; it comes from a compiled public write-up covering software, ML and forward-deployed roles.
03Architecture and ML / LLM design
Open discussion or a presentation of an architecture for a legal workflow; length not reported
The best-supported technical content. A 2026 ML engineer onsite asked how to train a language-model agent for long-context, long-horizon tasks, then how to design its reinforcement-learning environments and rewards. The compiled outline adds an architecture presentation for legal workflows such as retrieval over a long credit agreement where a clause depends on a defined term, reviewing thousands of contracts against a checklist, or choosing between long context and retrieval.
What it tests
- Retrieval over long, cross-referenced documents: chunking, definitions, structure
- Long context versus retrieval, argued with cost and latency numbers
- Agent design for multi-step legal work with tool calls
- Environment and reward design for training long-horizon agents
- Scoping a vague legal workflow into a concrete system
How to prepare
- Prepare one end-to-end design for a document-heavy legal task, including how citations are produced.
- Bring rough numbers for context length, KV-cache memory and prefill cost so you can defend long context versus retrieval.
- For agent questions, separate what the environment provides, what the reward checks, and how you stop reward hacking.
One ML engineer onsite report from 2026 describes an agent and RL design discussion; the legal-workflow architecture prompts come from a compiled public write-up.
04Evaluation and grounding
Discussion round on evaluation design, per the compiled outline; length not reported
The compiled outline lists a separate round on evaluation. Prompts cover tying every claim in an answer to a source passage and measuring how often claims are unsupported, building an eval set when there is no reference answer and expert time is scarce, and deciding what must happen before a new frontier model reaches customers.
What it tests
- Claim-level grounding and citation checking
- Measuring an unsupported-claim rate with automatic checks plus expert labels
- Building evals without ground truth on a fixed expert budget
- Safe model upgrades: regression suites, staged rollout, customer constraints
How to prepare
- Be ready to size an expert-labelling budget and explain how you sample what experts grade.
- Treat a missed or unsupported claim as costly in legal work, and say how your design reflects that.
- Do not trust public benchmarks alone; explain the internal checks a new model must pass.
No first-hand report describes an evaluation round; it and its prompts come from a compiled public write-up.
05Customer or partner scenario
Scenario discussion, per the compiled outline; length not reported
Listed in the compiled outline, mainly for forward-deployed and applied roles: you are given a realistic customer problem, such as a partner reporting that the product missed a change-of-control clause, and asked to debug it or estimate the cost and turnaround of a large review run.
What it tests
- Debugging a retrieval and LLM pipeline from its logs, stage by stage
- Estimating cost and turnaround for a large document review
- Communicating with a non-technical, demanding customer
How to prepare
- Work through parsing, chunking, retrieval, reranking and generation in order and say which log would show each failure.
- State assumptions out loud when estimating and name the first lever you would pull.
No first-hand report describes this round; it comes from a compiled public write-up and seems aimed at forward-deployed roles.
06Behavioral and project deep dive
In the one detailed onsite, about two of four interviews were behavioral, including a walk-through of a recent project
The one detailed candidate account describes an onsite that was mostly behavioral, with a discussion of a recent project as the main prompt. The company's careers page stresses simplicity, decisiveness and an intense, keep-improving culture, so expect questions on how you make calls and move fast.
What it tests
- Depth behind a recent project: decisions, trade-offs, results
- Bias for action and making decisions with incomplete information
- Customer focus and curiosity about legal work
- Fit with an intense, fast-moving culture
How to prepare
- Prepare one recent project in depth, including what failed and what you would change.
- Have examples of making a call quickly and correcting course afterwards.
- Read the company's careers page and recent posts so your answer to why Harvey is specific.
One ML engineer report from 2024 says the four-interview onsite was mostly behavioral, centred on discussing a recent project; no other first-hand account yet.
Questions and themes reported most often
- Walk through a recent project you worked on (1 reports, behavioral and project deep dive)
- Train a long-horizon agent with RL: design the environments and rewards (1 reports, architecture and ml / llm design)
- Retrieval over a long contract where a clause depends on a defined term (architecture and ml / llm design)
- Whole contract in context or retrieval? Defend it with numbers (architecture and ml / llm design)
- Chunk a legal document without splitting a clause (paired coding)
- Link every claim to a source passage and measure the unsupported-claim rate (evaluation and grounding)
- What must happen before a new frontier model reaches customers (evaluation and grounding)
- Debug a missed change-of-control clause in contract review (customer or partner scenario)
What Harvey asks about
- LLM / RAG system design35%
- Evaluation & grounding25%
- Behavioral & project deep dive20%
- Coding10%
- Agents & RL10%
Harvey interview prep checklist
- Asked the recruiter which rounds, formats and tools to expect
- Can sketch retrieval over a long contract that resolves defined terms and cites passages
- Can argue long context versus retrieval with rough cost and memory numbers
- Designed an eval plan with an unsupported-claim metric and an expert-labelling budget
- Wrote a chunker that never splits a unit and keeps overlap between chunks
- Prepared a deep dive on one recent project, including what did not work
- Have a specific answer to why Harvey and why legal AI
Tips
- Confirm the rounds with your recruiter early; first-hand accounts of this loop are scarce and the process may differ by team.
- Make grounded retrieval over long legal documents your main design topic: structure-aware chunking, defined terms, cross-references and citations.
- Prepare to design evaluations where expert time is expensive and an unsupported claim is worse than a missing one.
- Have numbers ready for long context versus retrieval: context length, KV-cache memory, prefill time and token cost.
- For agent roles, rehearse reward and environment design for long multi-step tasks, including how to verify outcomes.
- Practise paired coding on document-processing tasks rather than only classic algorithm problems.
- Pick one recent project and prepare a deep dive; in the one detailed account it was the main prompt.
- Show curiosity about how lawyers work; legal experience is not required but domain care is valued.
Practice the Harvey loop
Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.
Harvey interview FAQ
How many interview rounds does Harvey have for ML roles?
Candidates report 6 stages: Recruiter screen or direct outreach, Paired coding, Architecture and ML / LLM design, Evaluation and grounding, Customer or partner scenario and Behavioral and project deep dive. First contact may come from a recruiter or, in one 2024 report, directly from a founder who skipped the phone screen entirely. That candidate's onsite had four interviews, about half technical and the rest behavioral. A compiled 2026 write-up describes a longer sequence for software, ML, forward-deployed and research engineers: recruiter screen, a paired coding round, a presentation of an architecture for a legal workflow, a round on designing evaluations, and a customer or partner scenario. A separate 2026 ML engineer onsite included an open discussion of reinforcement learning for long-horizon agents. The company's own pages do not describe the interview process. This outline rests on two first-hand reports plus one compiled write-up, so expect the rounds to differ by team and level.
How long does the Harvey ML interview process take?
Reported timelines run Not reported.
What does Harvey ask in ML interviews?
By reported question mix: LLM / RAG system design 35%; Evaluation & grounding 25%; Behavioral & project deep dive 20%; Coding 10%; Agents & RL 10%. The most repeated themes: Walk through a recent project you worked on; Train a long-horizon agent with RL: design the environments and rewards; Retrieval over a long contract where a clause depends on a defined term.
How should I prepare for the Harvey ML interview?
Confirm the rounds with your recruiter early; first-hand accounts of this loop are scarce and the process may differ by team. Make grounded retrieval over long legal documents your main design topic: structure-aware chunking, defined terms, cross-references and citations. Prepare to design evaluations where expert time is expensive and an unsupported claim is worse than a missing one.
Are there free Harvey interview questions?
LeetML has 10 questions reported from Harvey 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
- GleanGraph-search coding, an AI-assisted build round and enterprise ML design, judged on a small set of 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.
- CohereEnterprise LLM lab: post-training and retrieval design, numpy ML coding, a research talk and, for some teams, AI-assisted coding.
- SierraAgent-building judgment: where guardrails live, how to evaluate a conversational agent, and how you work with customers.
- PerplexityPractical multi-part coding, retrieval and LLM depth, a hiring-manager project deep dive and a real check on product taste.