Harvey AI Interview Process & Rounds
Harvey's loop is unusually predictable, and the reports say so directly: twelve of the fifty threads in this corpus describe the question they got as 地里的老题 — the known one from the forum. The phone screen (店面/电面) is a single coding question drawn from a small rotating set, most often the citation/text-matching problem or the spreadsheet cell-and-formula problem, on CoderPad for about an hour. The onsite (昂赛) adds a system design round that is almost always either Google Drive-style file storage with access control or a RAG pipeline — RAG being, as one report puts it, 他们家的基本盘 — plus a project deep dive and a values-based behavioral round. The hiring-manager round is a full hour on one past project, and reports describe it as genuinely selective on domain fit rather than a formality.
Key facts
- •4 distinct round types
- •9 questions reconstructed from 85 candidate reports
- •Reports span Mar 2025 – Aug 2026
- •Refreshed monthly · last updated August 2026
The Harvey AI loop, from candidate reports
Harvey's loop is unusually predictable, and the reports say so directly: twelve of the fifty threads in this corpus describe the question they got as 地里的老题 — the known one from the forum. The phone screen (店面/电面) is a single coding question drawn from a small rotating set, most often the citation/text-matching problem or the spreadsheet cell-and-formula problem, on CoderPad for about an hour. The onsite (昂赛) adds a system design round that is almost always either Google Drive-style file storage with access control or a RAG pipeline — RAG being, as one report puts it, 他们家的基本盘 — plus a project deep dive and a values-based behavioral round. The hiring-manager round is a full hour on one past project, and reports describe it as genuinely selective on domain fit rather than a formality.
What does Harvey AI ask in each round?
Harvey AI interviews span 4distinct round types, shown below. Counts reflect distinct questions per round across the loops we’ve indexed.
- •On the file-storage design: separate the metadata store from the blob store early, then raise chunking for large files before the interviewer asks — passers volunteer it, failures wait to be prompted
- •On the file-storage design: cover organization- and group-level ACL, not just user-level sharing, and be ready to explain how a pre-signed URL is scoped so chunk ownership can be verified
- •On the RAG design: walk the whole pipeline — crawler, chunking, embedding, vector store, retrieval, generation, evaluation — and treat crawler scalability as a first-class dimension rather than an afterthought
- •On the behavioral round: prepare three to five tight STAR stories mapped to Harvey's published values in advance, and keep them short in the 30-minute format so you cover breadth
- •On the hiring-manager round: pick a project that is genuinely domain-relevant to the team, and lead with why and what you traded off rather than waiting to be probed
- •Read the prompt HR sends before the design round — reports say it is detailed enough to tailor preparation around
- •Over-indexing on LLM and RAG internals while neglecting web-crawler scalability — one candidate was strong in every other round and failed system design on exactly this
- •Designing only user-level ACL on the file-storage question, with no answer for organization-level sharing at scale
- •Spending so long on requirements clarification that the deep-dive portion gets squeezed — a repeated cause of running out of time
- •Treating the RAG question as a pure ML problem rather than a distributed-systems one
- •Bringing generic STAR stories that were never mapped to Harvey's specific values, or burning the 30-minute behavioral on a single story
- •Choosing a hiring-manager project that sounds relevant but has no depth under probing — reports describe the HM losing interest quickly
FAQ
How many rounds is the Harvey AI interview?▾
What is the Harvey AI interview process?▾
How fresh is this Harvey AI interview data?▾
Get the full Harvey AI catalog
Every question, every candidate-reported follow-up, and what passers actually do. Monthly refresh.