Poindexter Labs, an expert reasoning training data startup, has raised £2 million in an oversubscribed seed round led by Episode 1, with participation from Octopus Ventures' First Cheque Fund, notable angels, and a group of Poindexter's own contributors - the mathematicians and scientists who do the work - who invested personal savings into the platform they helped build. The raise comes as Poindexter signs its first direct contract with a major frontier AI lab, marking its transition from intermediary partner to named vendor.
The problem Poindexter was built to address is structural rather than incidental. Between 20% and 50% of AI training data is wrong, a figure researchers at frontier labs accept as a given. The major data platforms were designed for simple labelling tasks - tagging images, categorising text - and were never built for the step-by-step expert reasoning that next-generation models from OpenAI, DeepMind, Anthropic and others require. Those platforms isolate workers, use anonymous black-box review, and run adversarial processes that actively incentivise discarding tasks rather than improving them. The result is that platforms deliver only a fraction of contracted data and discard the rest, creating one of the most significant hidden constraints on AI progress.
Poindexter developed its methodology from the ground up with the people who do the work. Its contributors - Olympiad medalists, PhDs and professors from Oxbridge, Ivy League and MIT institutions, each admitted through a rigorous live subject knowledge interview - produce training datasets covering advanced mathematics, scientific reasoning, legal analysis, professional judgement, and domain expertise across medicine, finance and engineering. Their collaborative peer-review workflow, refined across millions of tasks, underpins a 95%+ data delivery rate and a 99.5% acceptance rate on completed work, against an industry average of 5% to 60%.
That workflow is now codified in Poindexter's proprietary platform, currently in beta. The platform operates in two modes: powering Poindexter's on-demand data production service for frontier AI labs; and available to licence directly to government departments and enterprises building their own AI systems, giving their internal experts the same collaborative tools Poindexter's contributors use. Founded in 2025 by Jocelyn D'Arcy - a former maths teacher and National Mathematics and Science College chief of staff with degrees from MIT, Cambridge and Oxford - Poindexter bootstrapped to $1.6 million in revenue in its first six months of operation before taking any outside funding. The company has an accepted ACL paper and a growing research pipeline.
Poindexter Labs plans to scale the team, deepen direct relationships with Tier 1 AI labs, and bring the platform to market for enterprises and government departments building their own AI systems.
The workflow that has defined the industry since its inception was designed for a factory, not for knowledge creation. As a result, a huge chunk of training data is discarded not because it is wrong, but because adversarial review processes actively incentivise discarding tasks rather than improving them. This is a workflow problem, and we built Poindexter the way academics build knowledge: collaboratively, transparently, with peer review at every step. We have an accepted ACL paper, a growing research pipeline, and a commitment to building the kind of high-quality sovereign datasets the UK's AI future depends on. Our proof point is our revenue. If we weren't 5x more efficient than the platforms we replaced, we simply wouldn't exist.
Poindexter has done something rare - built a business that is both technically differentiated and commercially validated from day one. The traction speaks for itself. We backed them because the problem is real, the solution is working, and Jocelyn is unlike any founder we have ever met.
