An applied research lab that teaches frontier AI how a cardiologist, a tax attorney, and a quant actually think - then sells the lesson plan back to the labs building the models.
She is being paid for it. So is the radiologist next to her. So, in a sense, are the next 99,998 specialists behind them.
If you walked into AfterQuery's office tomorrow, you would not find a server farm or a wall of GPUs. You would find a project manager arguing about a rubric. A lawyer on a Zoom call explaining why the model's contract redline is technically right and practically suicidal. An engineer staring at a benchmark, frowning.
This is what a data company looks like in 2026. Not labelers in a warehouse. Specialists at a screen, on contract, teaching the smartest software ever built how to do their jobs - and getting paid the kind of money that makes the moonlighting worth it.
AfterQuery is the company that built the teachers' lounge. Roughly 100,000 vetted experts on one side. Anthropic, OpenAI, and friends on the other. In the middle: datasets, benchmarks, reinforcement-learning environments, and the unglamorous machinery that turns a Wharton MBA's marginal hour into model weights.
Encyclopedic, opinionated, occasionally medically dangerous. Not great at explaining itself.
Around late 2024, a quiet panic set in among the labs. The scaling laws were holding. The compute was there. The text was not. Anything left on the open web was either already memorized, already poisoned, or written by a model trained on the prior round of itself - an AI ouroboros eating its own tail.
The way out, if there was one, ran through people. Specifically, the people who had spent a decade or three becoming undeniably good at one thing. Doctors. Litigators. Bond traders. Cracked software engineers. The catch: those people are expensive, hard to find, and almost universally not interested in labeling cat photos.
So the question on the table was simple, even if the answer was not. Could you build the back office for human expertise at AI-lab speed? A vetting funnel, a workflow, a rubric system, a payment pipeline - all of it - for tens of thousands of specialists at once?
A respectable resume to set on fire.
Spencer Mateega and Carlos Georgescu started AfterQuery in January 2025. Mateega had bounced through Silver Lake, Morgan Stanley, Meta, and Google before deciding that none of them were quite the experiment he wanted to run. A B.S. in finance and statistics from Wharton. A master's in computer science from Penn. The kind of credential stack that gets you a corner office and a mortgage. He chose YC instead.
The bet was unfashionable when they made it. Most of the data-labeling category had spent five years racing toward cheaper labor. AfterQuery's pitch was the opposite - go upmarket. Recruit the people whose hourly rates make CFOs wince. Build the tooling that makes their hour fifty times more productive. Sell the output to the only customers who could afford it: foundation-model labs racing each other through the trillions.
CEO. Wharton + Penn CS. Ex-Silver Lake, Morgan Stanley, Meta, Google. The one on the calls.
Co-founded the company out of YC Winter 2025. The one in the rubrics.
Everything else is plumbing.
Prompt-response pairs and step-by-step reasoning traces, written by people who actually do the job. Less "label the image," more "show your work."
Expert-designed prompts paired with grading frameworks. The model attempts, the rubric scores, the policy improves. Repeat ten million times.
Sandboxes wired to real APIs and tools so agents can be trained and evaluated on jobs that look like jobs - not toy chess.
Humans demonstrating the unglamorous reality of getting things done in software, click by click, for models that want to do the same.
Catalog page, photographed under fluorescent light. The labels are boring on purpose.
Take a moment. Read that sentence again.
Bar chart, hand-lettered axis. The middle bar is the one investors squinted at.
The shorter version: teach machines how experts think.
AfterQuery's public mission is to "encode domain-specific excellence into forms that machines can learn." The unpublished version is that the next leap in AI capability is locked inside specialists, and the lab that gets to them first wins the next benchmark.
That second version is, conveniently, also a business model.
The customers are the labs you would expect: frontier foundation-model developers and a small set of large enterprises building agentic systems. The expert side counts roughly 100,000 vetted specialists across finance, law, medicine, software, and a long tail of niches that get harder to staff every quarter.
Scale AI, Surge, Mercor, Invisible, Snorkel, Labelbox. All chasing the same gold rush from different angles. AfterQuery's angle is the rubric.
That somebody is currently a 120-person startup with a 14-month head start.
The interesting thing about AfterQuery is not the funding round. Funding rounds are weather. The interesting thing is the structural bet: that the most valuable input to a frontier model is no longer compute or tokens, but the captured judgment of people who have done the work.
If that bet is right, every serious AI capability over the next five years - in medicine, in law, in finance, in engineering - will, at some point, pass through a database of expert-graded examples. AfterQuery's, or someone's. They are betting on theirs.
Back to the office, the Tuesday morning. The analyst finishes grading the DCF and clicks submit. Somewhere across town, a model that has never met her becomes, in a way that is hard to measure but real, slightly better at her job. The rubric goes into the queue. The next prompt loads. The teacher's lounge stays open.
That is what AfterQuery built. Quietly, in fourteen months, with a hundred and twenty people and a list of phone numbers most labs would kill for.
Compiled from public reporting and the company's own pages. Numbers reflect disclosures as of April 2026. verified