Harvey interview, decoded.

Published accounts describe four to six stages over three to five weeks for engineering, with a founder round at senior. The rounds below are the engineering loop; the legal engineer track is described under the tracks and rests on Harvey's own postings and hiring post rather than candidate reports.

The reports agree on what is screened beyond the code: product judgement about legal work, accuracy under zero tolerance for a wrong source, and ownership at speed.

• Product judgement informed by how legal work actually runs: due diligence, drafting, research, review
• Grounded output: every answer cites a real source, and a fabricated one is a failure of the system
• Comfort with enterprise security: tenant isolation, encryption, privilege-aware access
• Ownership and speed in ambiguity, shipping with non-engineers involved
• Curiosity about the customer; Harvey's own hiring post says this is required of every role

The reported questions come from 49 candidate reports on PracHub and 17 on 1Point3Acres in 2026, and they cluster tightly: text with structure in the coding rounds, documents at scale in design.

Harvey has no timed online assessment; the filter is a 60-minute coding screen, then the onsite coding rounds. Forty-nine reported questions across 2026 share one shape, text under structure, and the same handful of problems recur across reports. The exact statement varies by interviewer, so what follows is each problem's family and the approach that solves it.

• A spreadsheet engine with cycle detection: Model cells as nodes and formula references as edges. Evaluate with a depth-first search that marks a cell visiting, then done; meeting a visiting cell is the cycle. For updates, keep the reverse dependency graph and re-evaluate dependents in topological order. Get evaluation working before propagation; the cycle case is what the follow-up asks for.
• Find duplicate files safely at scale: Group by size first, then by a hash of the first few kilobytes, then by a full hash, so almost no file is read in full. Stream the hash rather than loading the file; say out loud that a hash collision is possible and what a byte-compare fallback costs. The 'safely' is about never deleting on a hash alone.
• An in-memory file vault with duplicate-name handling: A tree of directory nodes, each with a map from name to child. On a name collision, generate the next free suffix ('memo (1).pdf') by counting from the map rather than scanning. Support path resolution, listing and move; a trie-shaped tree keeps every operation proportional to path depth.
• Highlight overlapping phrases in a sentence: Turn each match into an interval over character offsets, sort by start, and merge intervals that touch or overlap into one highlight. When highlights carry a source, merge equal spans and rank their citations by the count or weight behind them; it is interval merging with a tie-break.
• Retrieval-augmented answers with citations over a document vault: Chunk by structure (sections, clauses) rather than fixed size, embed, retrieve the top k, rerank with a cross-encoder, then generate with the passages in context and require each claim to point at a passage id. A grounding check that rejects a citation not in the retrieved set is the part interviewers wait for.
• A memo Q&A agent for a large law firm: The same retrieval loop, wrapped in an agent with tools (search, fetch a document, check permissions), a per-user permission filter applied before retrieval so a privileged memo never reaches the context, and an evaluation set of real questions with known sources. Name the failure you are designing against: a confident answer citing a memo the user cannot see.
• A secure PDF data room: Tenant isolation at the storage and query layer, encryption at rest with per-tenant keys, per-document access lists, watermarked viewing, an audit log of every open; then the product questions: bulk upload, indexing for search, redaction. Security first, then features, is the order that reads as understanding the customer.

Practise in Python with its idioms in hand (comprehensions, dictionaries, sets, collections), and practise talking while you type: every account describes interviewers who push on details to see how you take feedback. One 2026 report names an AI-assisted coding onsite; if your recruiter confirms tools are allowed, the round grades how you direct and verify the assistant, not whether you used it.

The medians are Levels.fyi's software engineer figures for the San Francisco Bay Area, L3 to L5, read on 21 September 2026 from a small sample; the posted ranges are Harvey's own on its careers pages the same day (a senior AI platform role and the law-firm corporate legal engineer role). London roles are reported at roughly 30 per cent below San Francisco base. One third-party guide estimates $300K to $450K total at mid, $450K to $650K at senior and $650K to $950K at staff; treat those as estimates. Equity has been marked up through four rounds in eighteen months, which is the thing to ask about: strike price, vesting, and whether the grant is priced at the $15.5 billion round.

1. 01Solve the spreadsheet problem before they ask it. Cells, formulas, dependency graph, cycle detection, then update propagation, in Python, under 45 minutes.
2. 02Have the retrieval design ready to draw, with the grounding check: chunk by structure, retrieve, rerank, generate with passage ids, reject any citation not in the retrieved set.
3. 03Read about how legal work runs. Every account says candidates who can picture a due diligence or a contract review make better product calls in the design and deep-dive rounds.
4. 04Prepare one project for the deep dive as if it were the behavioural round, because it is: ownership, ambiguity, and shipping with non-engineers in the room.
5. 05Legal engineers: rehearse a live demo of a legal use case for a sceptical partner, and know which of Harvey's three legal roles you are asking for before the recruiter call.
6. 06Ask the recruiter whether the coding onsite is AI-assisted; at least one 2026 loop was.

This guide covers Harvey's software and AI engineering loop and the legal engineer track, which is a role for qualified lawyers rather than developers. For management and leadership roles the loop is similar but the bar shifts to people, delivery and strategy, so pair it with the leadership interview prep hub. The bar for your exact role comes from the role-by-role guides, and the prep that actually transfers is spoken, so run a mock interview before the real one.

• 01PracHub, Harvey interview questions49 reported questions in 2026 (19 coding, 8 system design, 3 ML system design), the 'text under structure' pattern, and the named problems: duplicate files at scale, the in-memory file vault, overlapping highlights, merged highlights with ranked citations, RAG over a document vault, the memo Q&A agent, the secure PDF data room, the rule-guided evaluation agent.
• 021Point3Acres, Harvey interview reports17 reports from April to September 2026: the spreadsheet and Excel cell problems, duplicate file finder, Vault in-memory file directory, citation highlighting, design a text editor, design Vault, the memo Q&A agent, and one AI-assisted coding onsite.
• 03Aced (Exponent), interviewing at Harvey AI'Implement spreadsheet' and 'tell me about your past projects' as recently asked, and the tokenisation problem with simulated embeddings and no ML models.
• 04TechInterview, Harvey interview guide 2026the stages and durations (recruiter screen, 45-minute hiring manager screen, 60-minute coding screen, four-to-five-round onsite, founder interview at senior), three to five weeks, the round-by-round focus, and the third-party pay estimates by level; updated 3 July 2026.
• 05Aced (Exponent), Harvey ML operations engineer interviewthe paired coding on tokenisation and vector databases, the solution architecture presentation, the director round, four to six weeks, and that collaboration is judged through the technical work rather than a separate behavioural round.
• 06Taro, Harvey senior software engineer, August 2025first-hand: a difficult multi-part screening problem at LeetCode medium or above, syntax lookups allowed, a helpful interviewer, and the candidate's read that it seemed a stressful place to work.
• 07Harvey, The ultimate guide to landing a job at Harvey (Katie Burke, COO, 5 August 2026)the three legal roles (legal engineers, legal innovation partners, applied legal research), that legal engineers veer toward sales and customer success, and that every role needs curiosity about customers.
• 08Harvey careers, Legal Engineer (Law Firm, Corporate)JD or equivalent and 3+ years in practice, pilots and live demos, the initial recruiting interview covering all three legal roles, and $220,000 to $320,000 on-target 75/25 plus equity in San Francisco; read 21 September 2026.
• 09Harvey careers, Senior Software Engineer, AI Platform$220,000 to $300,000 plus equity, San Francisco office-based, and the line that legal is one of the hardest domains for AI with zero margin for error; read 21 September 2026.
• 10Levels.fyi, Harvey software engineer, San Francisco Bay Areathe L3 to L5 medians, about $390K to $492K total, read 21 September 2026.
• 11TechCrunch, Harvey hits a $15.5B valuationthe $550M round on 9 September 2026 co-led by Diffusion and Lightspeed, more than $1.55B raised in total, the $11B round in March and $8B in December 2025, and Harvey Tenet, its first in-house model.
• 12Harvey, raising at an $11 billion valuation (25 March 2026)the $200M round co-led by GIC and Sequoia, more than 100,000 lawyers across 1,300 organisations, a majority of the AmLaw 100, and the product names: Vault, agents, Spaces, Knowledge, Memory.
• 13Wikipedia, Harvey (software)the founders and their backgrounds, the founding in San Francisco in 2022, and the funding history from the OpenAI Startup Fund seed to the September 2026 round.

Interview processes change. This reflects widely-reported and sourced conditions as of 2026 — confirm specifics with your recruiter, and treat it as a map rather than a guarantee.