Mercor built an AI that interviews people and scores how well they will do a job. Apply for an engineering role there and you meet that product from the wrong side of the table: one of your rounds is run by their own bot, which reads your resume and improvises questions off the exact phrasing you used. Getting cross-examined by the thing you might end up building is a fair preview of the work.
Mercor was three founders in their early twenties not long ago, and it is now reportedly in talks to raise around $500 million at a $20 billion valuation, after running from roughly zero to a reported $500 million revenue run-rate in about 17 months. The engineering team is small against that scale, so the bar is generalist and the loop is quick. A few weeks from the first call to an offer is normal.
What Mercor does, and why the loop looks the way it does
Mercor is a marketplace that connects vetted domain experts to the AI labs that need human judgment. People who write code, argue law, read scans, or model deals get hired to produce training data, run reinforcement-learning feedback, and grade model output. More than 30,000 contractors sit on the supply side. The company has said it pays over $1.5 million a day to the humans training these models, at an average around $95 an hour.
The hard engineering problems fall straight out of that. How do you rank tens of thousands of specialists against an incoming contract and route the right ones to it? How do you check a candidate’s expertise without putting a human in the loop on every application? How do you build one interviewer that scores fairly across contract law and CUDA kernels? When the interview leans on ranking, matching, and product judgment under ambiguity, it is testing the actual job rather than a stand-in for it.
The pace matters too. Brendan Foody (CEO), Adarsh Hiremath (CTO), and Surya Midha (COO) started Mercor in their early twenties, and the valuation ran from $2 billion in early 2025 to $10 billion by that October on a $350 million Series C. A company moving that fast interviews for people who ship and defend decisions, not people who wait for a spec.
The loop, from recruiter call to super day
The process is short and broad: a recruiter call, an interview run by Mercor’s own AI, a technical phone screen with an engineer, and a virtual super day that eats most of a morning. It can feel uneven, because one loop is trying to read both raw coding ability and product instinct.
| Round | Format and length | Run by | What it is screening for |
|---|---|---|---|
| Recruiter screen | 20 to 30 minute call | Recruiter | Motivation, background, comp expectations, timeline |
| AI interview | Automated, resume-driven, questions generated on the fly | Mercor’s own AI interviewer | Whether your resume claims survive specific, reactive follow-ups |
| Technical phone screen | 45 to 60 minutes, live coding | An engineer | Practical coding, reasoning out loud, correctness under a clock |
| Virtual super day | About 5 hours, back-to-back sessions | Engineers and a hiring manager | System design, a longer coding build, AI-assisted vibe coding, product judgment |
The AI interview is where most people underprep
Candidates expect a canned video screen and get something reactive instead. The bot ingests your resume and generates follow-ups from the specific claims on it. Say you ‘scaled the matching service’ and it will ask what the bottleneck actually was, which number moved, and what you would change now. Loose lines that survived a human skim get pulled apart, because the model has nothing better to do than keep tugging the thread you handed it.
Prep the way you would for a sharp staff engineer who has already read your resume twice. Know every bullet cold, with the real number behind it, the real failure you hit, and the tradeoff you would revisit. Cut anything you cannot defend for five minutes. Speak in specifics, because the system follows your wording, and a generic answer only hands it more surface to probe.
The engineer phone screen, the one conventional round
The human phone screen is a straightforward live-coding hour, and it is the most ordinary part of the loop. The problems stay practical: manipulate strings and hashes, handle the edge cases in a data-cleaning function, get something correct and readable inside the window. Reviewers care more about how you narrate your thinking and recover from a wrong turn than about whether you land the optimal solution on the first try. Talk through your assumptions before you start typing.
The super day, and the vibe-coding round nobody preps for
The onsite runs roughly five hours of back-to-back sessions. The coding stays practical rather than pure algorithm trivia. Expect to parse messy real-world input, dedupe a stream of candidate records, or wire up an endpoint under a clock, reasoning out loud the whole way. Correctness and clear thinking beat a clever one-liner here.
The system design round tends to map straight onto the business. Versions that come up:
- Design the pipeline that takes a resume, runs an automated interview, and returns a hireability score.
- Design the ranking system that matches 30,000 contractors to an incoming labeling contract and routes work to the best fits.
- Design a check that catches contractors gaming an AI-graded task.
Then there is the part people are not ready for: a vibe-coding session where you build a real feature with the AI tools you would actually reach for, Cursor or Copilot included. What they are watching is whether you can steer a model toward working software and then stand behind every line it wrote. Ship something that runs. When the model slips in a bug, catch it, and be ready to explain the tradeoff you accepted. Treating the assistant as an oracle you never audit is the quickest way to lose this round.
Reading an offer here
Cash bases land in the range you would expect from a well-funded San Francisco startup fighting the labs for the same engineers, but the exact figures move and are worth confirming on levels.fyi and in writing before you anchor on anything. A number from a forum thread is not your offer.
Equity is the real conversation, and at a company that repriced from $2 billion to $10 billion in eight months it is also the real uncertainty. Ask for the share count, the strike price, the latest preferred price, and the vesting schedule, then do the dilution math yourself. Given how often the valuation has jumped, ask whether earlier rounds carried any tender or secondary, because paper worth $20 billion only helps if there is a path to selling some of it. Push for the numbers that let you compare offers, not the headline valuation.
If you have a week
Spend the first evening rewriting your resume so every line carries a number and a decision you can defend, since both the AI round and the human rounds attack it. Run a couple of practical coding sessions against a clock, the kind where you clean ugly input rather than invert a binary tree. Sketch two or three ranking-or-matching designs end to end, because that shape keeps recurring. And build something small with Cursor before the super day, watching for the spots where the model steers you wrong, so the vibe-coding round is not the first time you catch it.
The through-line is judgment under speed. Mercor is betting that a small team can out-execute much larger ones, and the loop is built to find people who make good calls on incomplete information and own them afterward. If you need a problem fully specified before you move, five hours of this will make that plain to everyone in the room.
Practice the behavioral round:
