Exa doesn’t close its process with a whiteboard problem. It flies finalists to San Francisco to sit with the team and build something real for a day or two, and that working session is where the offer decision actually gets made. Everything before it exists to earn you the plane ticket.
If you haven’t run into Exa yet, it’s the company building search for AI agents rather than for people. The API takes a query from a model and returns clean, structured results in milliseconds, backed by custom crawlers that track more than 500 billion URLs and embedding models trained on an in-house GPU cluster. It started in 2021 as Metaphor, founded by Will Bryk and Dan McArdle, and raised a $250M Series C led by Andreessen Horowitz in May 2026 at a $2.2B valuation, roughly triple where the Series B landed eight months earlier. Around 120 people work there. Cursor, Cognition, OpenRouter, and HubSpot sit among the 5,000-odd companies and 400,000 developers calling the API.
That context shapes the interview. A company this small, growing this fast, screens hard for two things: the quality bar on a genuinely difficult technical problem, and a culture the founders describe in unusually blunt terms. Bryk has said on record that he wants “rebellious” engineers, people who don’t accept that search has to work the way Google decided it works. He’s also said experience is “not as important” as whether you actually care about building search or large-scale systems. Read those two statements together and the rest of the process falls into place.
The shape of the loop
The stages below are what candidates and Exa’s own hiring notes describe. Exact naming shifts by team and by whether you’re going in as a new grad or a staff hire, but the arc is stable: a values-heavy first call, a technical screen, then the in-person build that carries most of the weight.
| Stage | Format | What it’s really testing | Typical length |
|---|---|---|---|
| Intro call | Video with a founder or hiring manager | Motivation, values, whether you actually use and think about search | 30 to 45 min |
| Technical screen | Live coding and systems talk in a shared editor | Writing real code fast, reasoning about retrieval and latency | 60 min |
| Take-home or paired build (role-dependent) | A small project against the API or a retrieval problem | Clean code and sound judgment without someone watching | A few hours to a few days |
| On-site working trial | One to two days building with the team in San Francisco | Whether you can ship with these people and handle ambiguity live | 1 to 2 days |
| Final and offer | Founder conversation plus reference checks | Long-term conviction and values fit | 30 to 60 min |
The first call is a values screen wearing a technical hat
The opener is usually with a founder or an engineer who would be close to you day to day. Do not treat it as a formality. Exa has been open that it weighs values above the resume, so the questions probe why you want to work on search specifically rather than at whichever AI company is hottest this quarter. A weak answer here reads as “I applied to twenty labs” and quietly ends the process.
Come in having actually used the product. Sign up, run real queries through the API, compare a few results against a plain web search, and form an opinion about where it’s sharp and where it’s rough. Saying “I ran your neural search on a research question and it surfaced three papers Google buried under SEO junk, but the highlights felt off on long PDFs” tells them more than any credential. That is the rebellious-but-grounded signal they’re listening for.
The technical screen: code plus retrieval reasoning
The live screen is a shared-editor coding session with a discussion layered on top. The coding half is standard for a strong startup: a data-structures or string problem you’re expected to write cleanly and talk through, not a puzzle designed to trip you. Think along the lines of parsing and deduplicating a stream of URLs, merging ranked result lists, or maintaining a top-k set under a memory bound. What separates candidates is whether you write code that would survive review and whether you catch the edge cases before the interviewer points at them.
The discussion half moves into retrieval and systems. Expect questions phrased close to how the work really sounds:
- How would you find the nearest vectors to a query embedding across a billion documents without scanning all of them?
- What breaks first when you push an approximate-nearest-neighbor index from one machine to a sharded cluster?
- A crawler is hitting the same domains too hard and missing fresh pages elsewhere. How do you schedule it?
- Where does latency actually go in a search request, and which part would you attack first?
You are not expected to have built a web-scale index. You are expected to reason from fundamentals: why HNSW trades memory for recall, why IVF partitions the space, what recall-versus-latency knob you’d turn for an agent that needs an answer in under 300 milliseconds versus a batch job that can wait. If you can sketch a rough plan and name your assumptions out loud, you’re doing the round right.
The working trial, and why it decides everything
This is the round that makes Exa unusual, and it’s the one worth preparing for hardest. Finalists come to San Francisco and work with the team for one to two days on something close to real. Sometimes that’s a scoped feature against the codebase, sometimes a retrieval or evaluation problem the team genuinely cares about. You’ll pair, ask questions, read unfamiliar code, and ship an increment by the end.
The trial exists because a 45-minute algorithm round tells you almost nothing about whether someone is good to build with. In two days the team learns how you handle being dropped into a large codebase, whether you ask sharp questions or flail silently, how you react when your first approach doesn’t work, and whether you push back on a bad idea or just nod. That last one is the rebellious trait made observable. They want to watch you disagree well.
How to show up strong: move, don’t stall. Read enough of the surrounding code to be safe, then start writing and course-correct out loud. Ask the specific question rather than the vague one. When you hit an assumption, name it and check it instead of building on top of it for three hours. Ship something small and working over something ambitious and broken. And treat the engineers around you as future teammates, because for a day or two that is exactly what they are.
Retrieval and systems depth they actually probe
Because search is the whole company, the technical bar clusters around a few areas more than a generic backend loop would. Embeddings and vector search come up constantly: how they’re trained, why cosine similarity, when a reranker earns its cost. Crawling and freshness matter, since a stale index is a broken product. Latency budgets are treated as a first-class concern, not an afterthought, because agents calling the API in a loop feel every millisecond. And evaluation is a real topic, because “is this result good” is genuinely hard to measure at scale and the team lives in that ambiguity.
If you’re coming from a pure application-engineering background, the fastest way to close the gap is to build a tiny retrieval system yourself before the loop. Embed a few thousand documents, stand up an ANN index, add a reranking pass, and measure recall against a handful of hand-labeled queries. You’ll have opinions after an afternoon of that, and opinions grounded in something you built are exactly the currency this interview trades in.
Reading a Series C offer without fooling yourself
Compensation at a company this stage is base plus equity, and the equity is where the story is. Bay Area engineering salaries at a well-funded Series C startup tend to land in a broad band that Levels.fyi and the recruiter can pin down for your level far better than any static number here, so ask directly for the range rather than guessing. The part people misjudge is the stock.
To value an Exa grant, get four numbers from the recruiter: how many options or units you’re granted, the strike price, the most recent preferred share price from the Series C, and the total shares outstanding. That last one turns your grant into an actual ownership percentage, which is the only figure that means anything. A grant quoted as a large dollar amount at the latest preferred price is a real possibility, not a guarantee, and it stays illiquid until an exit. Weigh it against the near-certain cash a public company would hand you, and decide how much you believe in the bet. If you think search for agents is a durable category and Exa wins it, the math gets interesting. If you don’t, no equity number should talk you into it.
How to prepare in the two weeks before
Use the product until you have opinions about it. Refresh your data structures and write real code in a shared editor with a friend watching, since the live rounds reward fluency over cleverness. Read one solid explainer on approximate nearest neighbor search and build the small retrieval demo described above so retrieval questions feel concrete instead of theoretical. And prepare a real answer to why search, why Exa, why now, because the founders ask versions of it at every stage and a canned response is easy to spot.
The engineers who do well here show up already caring about the problem, and they spend the working trial acting like they already have the job. Exa is trying to find people it wants to sit next to for years. The best way through the process is to be visibly one of them.
Practice the behavioral round:
