Confirm you’re prepping for the right thing first. AfterQuery is the Y Combinator W25 company building what it calls an applied-research lab for data: it pays a large network of domain experts to document how a professional reasons through a task, so a model can train or be graded on real judgment rather than scraped text. If your search is turning up SQL triggers and result-set handling, that’s the wrong “after query.”
The raise, and why it means they’re hiring
On September 1, 2026, TechCrunch reported, picking up a Forbes story by Anna Tong, that AfterQuery had raised a round valuing it at about $3.2 billion. Read that carefully: it’s a press report, the company has not announced the round, and the lead investor and round size weren’t disclosed. YC partner Gustaf Alströmer is quoted calling it the fastest any YC company has gone from launch to unicorn, about 18 months out of the Winter 2025 batch.
The verified milestone is earlier and smaller. In its April 9, 2026 announcement, AfterQuery said it had closed a $30 million Series A at a $300 million valuation and had passed a $100 million annualized revenue run rate. Treat the $100M figure as a company statement, not an audited number. TechCrunch’s September piece names Nvidia, the legal-AI company Legora, and the Korean lab Motif Technologies as customers. A tenfold jump in months, real or reported, funds headcount, and the open board bears that out.
What the open roles tell you they screen for
Read the live postings before anything else. They carry titles, pay ranges, and a years-of-experience line (ranging from 1+ to 6+ years depending on role), but no skills list and no named technologies, so what follows is read from the titles and the product, not copied off a job description. The roles split into two tracks. The engineering track is the one most readers of this guide want: a Senior Software Engineer on Infrastructure & Platform, a Software Engineer on Applied AI / Platform, and a Software Engineer on “Hiring Infra” for the expert network. The second track, labeled Strategic Projects and Technical Strategic Projects, is closer to a forward-deployed or program role that owns a lab customer’s dataset end to end.
What these roles build is a two-sided platform: on one side, the systems that onboard, vet, task, and pay a network the company and press put at roughly 100,000 professionals; on the other, the pipelines that turn their output into clean, graded datasets a lab will pay for. So the screen is less “invert this binary tree” and more “can you build and run a data platform where the raw material is human judgment.” If you’ve shipped anything involving task queues, review workflows, payments to contributors, or large data pipelines, lead with it. The Applied AI title points toward model-adjacent work: running and grading model outputs, building evals, wiring LLM calls into the pipeline. AfterQuery sits in the same expert-data market as Mercor, Scale, and Surge, so the AI-native company interview guides hub and the AI-startup interview difficulty index help you calibrate the bar against its neighbors.
The interview loop, inferred from the postings and the product
Here the evidence runs thin: there is no reliable public account of AfterQuery’s engineering loop, and the few scattered Glassdoor entries sit across different roles and are too few to reconstruct one from. So the stages below are the shape an early-stage, fast-moving YC company of this kind usually runs, read against the role titles and what the product needs built. Nothing here is attributed to candidate reports, because there aren’t enough to attribute to.
- A recruiter or founder screen on background fit and why expert-data specifically. Have a crisp reason you want training-data infrastructure rather than “an AI startup” in general; at a company this small, a founder may take this call directly.
- A technical screen on working coding ability (most likely Python, though no posting names it), data handling, and clean problem decomposition. Expect a practical scripting or data-manipulation task over a hard algorithms puzzle, but keep both warm.
- A systems round designing a piece of the two-sided data platform: task routing, review state, ingestion. Prepare the design problem below, not a generic textbook prompt.
- For the Applied AI roles, a data-quality round on how you’d grade model outputs and measure dataset quality.
- A work-sample or paid trial on a real, scoped task, common at YC-stage companies; ask the recruiter whether a trial is part of the process.
- A founder or values round on ownership, speed, and judgment under a thin spec. Have one story about owning something ambiguous end to end and shipping it.
Expect it to move fast and to be less structured than a big-company loop: fewer rounds, but more weight on each one and on the founder’s read. A stage added or dropped mid-process is the stage the company is in, not a red flag.
The design problem AfterQuery lives
No posting spells this out, so prepare it as the problem the product is rather than a question you’ve been told to expect. Imagine you own the platform that turns that network of part-time domain experts into shippable datasets. A lab asks for 50,000 examples of how a tax attorney reasons through an ambiguous filing, each with the steps shown and a defensible answer. Walk the path end to end: how you route each task to a qualified expert and not a random one, how you hold review state as an item moves from drafted to peer-reviewed to accepted, how you detect and remove a contributor quietly using a chatbot instead of their own expertise, and how you pay everyone correctly across dozens of countries.
Take routing, where a strong answer separates itself. The simple design routes each task to whoever holds the matching qualification tag: a tax-attorney task goes to people tagged as tax attorneys. That’s easy to reason about and to defend: you can always show why a given expert got a task. But tag-only routing stalls the moment your best-qualified people are offline or saturated, and a 50,000-item order has a deadline. Routing by live availability keeps throughput up, but hands premium-rate work to whoever is merely adequate and online, surfacing as quality drift you won’t catch until review. The move most teams land on is qualification as a hard filter and availability as the ranking inside it: never route below the bar, and among those at or above it, prefer whoever can start now, with the winning tag logged so a disputed item stays traceable. Standard system design fundamentals carry the rest: queues, idempotency, a review state machine, and an audit log the lab can trust. What separates a strong answer from a component list is quality control: how do you know the dataset got worse after a rubric change?
The data-quality round, if you get the Applied AI track
For the model-adjacent roles, the hard questions are about measurement, not model architecture. Be ready to design a grading rubric for a subjective task and defend why two independent experts would score the same item the same way. Name the metric: inter-rater agreement between graders, as Cohen’s kappa for two or Fleiss’ kappa for more, reported corrected for chance rather than as raw percent-match. Track that separately from agreement against a held-out gold set the graders never see, which measures accuracy against a reference rather than consensus among peers. The regression check is re-scoring that gold slice after every rubric change and every change to the grader pool, alerting when either number drops below last week’s, so the canary catches a dataset getting worse before it ships. Be ready too on calibrating graders who drift and on using an LLM as a first-pass grader without its blind spots leaking in. Much of this rests on data plumbing, so brush up on the kind of SQL and query work you’d use to slice and audit a dataset, and keep a coding patterns pass warm so a scripting task is about judgment, not syntax. The premise here is that human judgment is the scarce input, which is why the AI-era interview guide‘s point lands specifically: showing judgment about model output is the thing being graded.
Compensation: read the posted ranges, and ask about equity
The YC jobs page lists salary ranges for AfterQuery’s roles; California pay-transparency law requires the posted ranges for its San Francisco roles, so these are the employer’s own numbers, not a crowdsourced estimate. As of November 2026:
| Role | Posted range (AfterQuery, YC jobs page, Nov 2026) | Level and location |
|---|---|---|
| Senior Software Engineer, Infrastructure & Platform | $220K–$280K | 6+ years; San Francisco |
| Software Engineer, Applied AI / Platform | $140K–$200K | 1+ years; San Francisco |
| Technical Strategic Projects (Lead and Associate) | $200K–$400K | spans levels; San Francisco |
| Software Engineer, Hiring Infra | $7K–$10K / month | 3+ years; remote, Latin America |
The San Francisco engineering roles show no remote or hybrid option, so read them as on-site; the only remote band for a full-time engineering role is the Latin America one, which pays against its own market, and the engineering intern posting also lists remote US. If you need visa sponsorship, the postings don’t say either way, so ask on the first call: at this headcount a company rarely has the immigration machinery in place, so the realistic default is “unlikely” until a recruiter tells you otherwise. There is one Hyderabad Glassdoor entry and no advertised India office, so treat an India-based role as unconfirmed rather than a posted path.
The gap in every posting is equity. No listing shows an options grant or range, and at a company whose reported valuation has moved this fast, the equity conversation matters more than the base. Ask the recruiter directly what the instrument is, how many shares or what percentage, the strike price, and the current preferred price, then model the whole package with the total comp calculator. A fresh, unconfirmed valuation makes a paper grant look larger than its realistic value, so push on base and grant separately in any salary negotiation, anchored on the posted range for your exact role.
The founder round and what they’re buying
A founder is likely in your final loop, and the signal they want is someone who ships under a thin spec. Because labs pay a premium for data that’s correct and defensible, have ready a story where you owned something ambiguous, set your own bar for “good enough,” and were right about it. Structure it with the STAR method if that helps; what they’re testing is whether you’ll move fast without shipping something wrong, because at a data company, wrong is the one thing the customer will notice.
Practice the behavioral round:
