Models ran out of internet before they ran out of appetite
Pretraining consumed the readily available text. What decides how good a model is now happens after that, and it needs people who actually know the answer.
A model that has read everything still has no way of knowing which of two plausible answers a specialist would accept. That judgment has to be supplied from outside, by someone qualified to make it, and it has to be supplied in a form the training process can use — a preference between two responses, a score against a written rubric, a corrected reference answer, an adversarial prompt that breaks a safeguard.
This is post-training, and it covers a family of methods: supervised fine-tuning on expert-written demonstrations, reward models trained on ranked comparisons and then optimised against with reinforcement learning, and direct preference optimisation, which skips the separate reward model and tunes on the preference pairs themselves. All of them share one input. Someone has to decide which output is better, and be right.
For general-knowledge tasks a careful generalist is enough. For a differential diagnosis, a securities filing, a proof, or a production incident, they are not. The gap between a crowdworker's label and a specialist's label on the same item is where model quality now moves, and it is why a board-certified cardiologist grading clinical reasoning commands a rate an order of magnitude above generalist annotation.
A two-sided market with a quality problem in the middle
Labs do not want a labour pool. They want a delivered dataset that holds up under audit. Everything between those two things is what Cortext is.
Scoping with the lab
Matching from the network
Delivery under review
6.8% of applicants join the network
The funnel is deliberately narrow, and every stage past the first is paid. Here is what each one filters for and what it costs to get through it.
Six positions, and what each one costs us
Operating principles are only worth publishing if they constrain something. Each of these does.
Rates are published before you apply
Unpaid screening is capped at 30 minutes
Rejected work gets a written reason
A human decides anything that ends work
No exclusivity, no non-competes
Rates are set by scarcity, not by your postcode
Who runs Cortext
The board seats five: two founders, one seat each for Halyard Partners and Meridian Growth Partners, and one independent director.
Ran the human-data organisation for a speech and dialogue research group before starting Cortext, where the recurring problem was that annotation vendors could supply volume but not judgement. Spent four years before that in clinical operations at a hospital network, which is where the conviction that credentialed reviewers produce categorically different labels came from. Holds a PhD in computational linguistics and still personally reviews the calibration report for every clinical project Cortext takes on.
Built and maintained the reward-model training pipeline for a mid-sized lab, including the tooling that flagged annotator drift mid-run. Owns Cortext's matching system, the task-routing service and the review infrastructure that scores every submitted item. Writes the quarterly note explaining, in public, which internal quality metrics moved and which did not.
Scaled a 9,000-person distributed operations function across eleven countries before joining Cortext in 2025. Responsible for project delivery, the regional hubs and the escalation path when a lab's timeline and an expert's availability disagree. Introduced the rule that no project ships to a lab until its rubric has cleared a human calibration set, which added roughly nine days to first delivery and cut post-delivery rework by a third.
Twelve years in finance operations at payments and marketplace companies, most recently running treasury for a cross-border payouts business covering 90 countries. At Cortext, owns the weekly payout run, the contractor tax reporting stack and the decision to absorb foreign-exchange spread rather than pass it to contributors. Reports payout failure rates to the board every month alongside revenue.
Joined from a professional licensing body, where the day job was verifying credentials at scale and adjudicating disputed ones. Runs recruiting, credential verification, the assessment bank and expert support. Wrote the thirty-minute ceiling on unpaid assessments into policy after auditing how long applicants were actually spending before anyone paid them anything.
Spent nine years as a research methodologist designing inter-rater reliability studies for medical and legal evidence review, and published on adjudication protocols for low-agreement domains. At Cortext, owns rubric standards, gold-set construction, reviewer certification and the appeals process for removed hours. The only executive with authority to halt a project mid-delivery on quality grounds.
Practised employment and commercial law for eleven years, including four advising platform companies on contractor classification across US, UK, EU and Indian regimes. Owns the contractor agreement, the confidentiality and IP terms labs require, data protection and the sanctions screening that determines which countries Cortext can pay. Maintains the position that Cortext does not ask contributors for non-compete or exclusivity terms.
$147.2M raised across three rounds
Cortext is venture-funded. Contributors are paid from revenue, not from the balance sheet — payouts do not depend on the next round closing.
Raised three months after incorporation, against a network of 400 experts and two pilot contracts.
Funded the credential-verification stack, the Nairobi and Bengaluru hubs, and the move from flat annotation work to rubric authoring.
Post-money valuation of $1.4B. Earmarked for clinical and legal expert recruitment, the agentic-trajectory tooling and a second payout rail.
One headquarters, four regional hubs, and a network everywhere else
The hubs exist to recruit and support experts in their regions and to keep working hours overlapping with the labs we deliver to. Contributors are never required to be near one.
San Francisco, CA 94104
United States
Media enquiries and company boilerplate
Journalists on deadline should write to press@cortext-ai.uk — we answer press mail ahead of everything else.
Cortext Labs, Inc. operates an expert marketplace for AI training data. Domain professionals — physicians, attorneys, engineers, quantitative analysts, linguists and research scientists — evaluate model output, author reference answers, build and apply grading rubrics, red-team frontier systems and annotate agent trajectories for the labs building large language models. Founded in 2024 and headquartered in San Francisco, Cortext works with more than 40 AI labs and research groups and pays its contributor network weekly in US dollars.