NuMind Turns Your Worst Documents Into Clean Data

The most tedious job in any office is reading a document just to copy three numbers out of it. An invoice total. A patient's diagnosis code. A clause buried on page nine of a contract. Humans hate this work, and generic AI models are surprisingly bad at it - they will happily invent a field that was never there. NuMind, a Paris startup that went through Y Combinator's Summer 2022 batch, is built entirely around fixing that one narrow, unglamorous problem.

The company's tagline is blunt: "Frontier AI for Document Extraction." That is the whole pitch. NuMind builds multimodal large language models that take a messy document - a PDF, a scan, a spreadsheet, a photo of a handwritten form - and return clean, structured JSON. You tell it what fields you want. It fills them in.

01 / What it doesExtraction, and only extraction

Most AI companies in 2026 are racing to build assistants that do everything. NuMind went the other way. It does one thing: turn documents into structured data. The interface is a JSON template. You describe the shape of the answer you want, hand over the document, and the model returns that exact shape - populated, typed, and ready to drop into a database.

A quietly important design choice sits underneath: the models are trained to leave a field blank when the answer is not in the document, rather than guess. In extraction work, a confident wrong answer is worse than no answer at all. NuMind's original model was fine-tuned on a synthetic dataset built specifically to discourage that kind of hallucination.

The mechanics stay deliberately simple. The original NuExtract took two inputs - a passage of text, up to a couple of thousand tokens, and the JSON template describing what to pull out - and returned the filled template. There is no prompt engineering to master, no chain of instructions to tune. If you can describe the fields you want, you can use the model. That plainness is a feature: it is what lets a developer wire extraction into a pipeline in an afternoon instead of a quarter.

02 / The peopleFrom Wolfram to Y Combinator

NuMind was founded in 2022 by Etienne Bernard and Samuel Bernard. Before starting the company, Etienne - now CEO - was head of Machine Learning at Wolfram Research, the company behind Mathematica and Wolfram Alpha. That background shows in the company's instinct to treat extraction as a modeling problem worth solving properly rather than a prompt to paper over.

The team is small - around 13 people - split between Paris, where it is headquartered at 5 Parvis Alan Turing, and a presence near Cambridge, Massachusetts. In March 2023, the company closed a $3 million seed round led by Flybridge, with Big Bets writing the first institutional check and Carya, Pioneer, Velocity, Sharpstone, Y Combinator and a group of angels joining in.

03 / The proof30 percent to 95 percent

The clearest evidence NuMind offers is a single client story. A customer was processing complex Japanese health-claim documents with an existing solution that got roughly 30% of the fields right. Running NuMind's extraction model, that number climbed to about 95%. The gap between those two figures is the whole business case: a general-purpose tool that mostly fails versus a specialized one that mostly works.

04 / ProductsOpen models, then a platform on top

NuMind's product line reads like a stack. At the bottom are open-source models it publishes on Hugging Face under MIT and Apache 2.0 licenses - no strings attached. NuExtract was the original text-to-JSON foundation model, fine-tuned from Microsoft's small phi-3-mini. NuNER and its zero-shot cousin NuNER Zero handle named-entity recognition. NuMarkdown converts documents into clean Markdown as a pipeline step.

NuExtract 2.0 is the current generation, and it added a sense the earlier models lacked: sight. Built on the Qwen-VL family, it is a vision-language model that reads page layout, tables and images rather than just a stream of text. It ships in 2B, 4B and 8B sizes, handles multiple languages, and is small enough to run on modest hardware. On top of these models sits the commercial NuExtract Platform - a hosted API that turns documents into JSON in any language, available either as a SaaS service or deployed inside a customer's own walls.

05 / The moatPrivacy is the feature

The documents worth extracting are usually the ones you cannot send to a stranger: medical claims, signed contracts, financial statements. That is why NuMind leans hard on private deployment. Because its models are compact and open, a hospital or law firm can run them on its own servers, and no sensitive page ever leaves the building. For regulated industries, that is not a nice-to-have - it is the reason a purchase is even possible.

It also explains the range of customers. NuMind's users span healthcare and health insurance, legal, finance and even scientific research - one recurring use case is parsing chemical reactions out of research papers. The common thread is not an industry. It is a document that a person would rather not read by hand.

When the company announced its seed round, it counted roughly a dozen private-beta customers already at work: sentiment analysis, job classification, legal document review. The list is telling less for its size than its spread. These are not variations on a single template - they are genuinely different problems that happen to share one shape: text in, structure out. NuMind's argument is that a model built for that shape, rather than for conversation, is the thing worth owning.

06 / The marketNarrow, on purpose

Document AI is crowded. Cloud giants offer OCR-plus-extraction services, a wave of intelligent-document-processing vendors chase invoices and forms, and any general LLM can be coaxed into extraction with enough prompting. NuMind's answer is not to be broader. It is to be better at the one thing, with models built for extraction and deployable where the data lives.

07 / The businessSmall team, big surface area

The model is open-core. Free, permissively licensed models pull developers in; usage-based API pricing and paid enterprise deployments pull revenue out; custom-model work handles the hardest client-specific cases. For a 13-person company, that is a wide surface to cover - which is precisely why the product stays narrow. Doing one job well is what lets a team this size compete with hyperscalers.

It is an old lesson dressed in new tooling. The founder's own framing - "most NLP tasks are unique" - is really an argument against the one-model-fits-all pitch. The generic model is a starting point. The value is in the last mile of accuracy on documents that look like nobody else's. NuMind sells that last mile.

The trajectory of the products tracks that thinking. Each release has narrowed the gap between a raw document and a database row: text-to-JSON first, then entity recognition and Markdown conversion, then a vision-language model that no longer needs a separate OCR step to read a scanned page. The platform, meanwhile, has moved toward a newer generation powered by a NuExtract 2.0 PRO model aimed at production reliability. The shape of the roadmap is consistent - fewer steps, cleaner output, the same job done better.

None of this guarantees the outcome. NuMind is early, small, and competing in a category where the largest technology companies on earth also have a product. But it has picked a real, boring, valuable problem and refused to be distracted from it. In a market full of AI that promises everything, a tool that reliably does one thing is a genuinely different bet.