Etienne Bernard Is Teaching AI to Read the Fine Print

Etienne Bernard likes a machine that knows when to keep quiet. In document extraction, the useful answer is sometimes a name, a date, or a total. Sometimes it is null, the small JSON admission that the requested fact simply is not there. NuMind, the company Bernard co-founded and leads, has trained its extraction models to make that distinction. It is a modest detail with expensive consequences: an empty field can be reviewed; a confident invention can travel surprisingly far through a company's systems before anyone notices.

This preference for disciplined automation runs through Bernard's career. He began in statistical physics, worked on faster ways to simulate dense systems, moved into machine learning at Wolfram Research, wrote a book designed to make the field approachable, and then built a company around language models that shoulder more of the user's burden. The objects changed from particles to predictions to PDFs. His taste did not. The technology should be sophisticated underneath and legible on top.

The Netflix Prize and the physicist

Bernard dates his first encounter with the phrase “machine learning” to around 2009, when the Netflix Prize was demonstrating the commercial force of a better recommendation algorithm. He was then immersed in statistical physics at École normale supérieure in Paris. The notion that a machine could learn struck him as both fascinating and powerful. He decided to dive in, as he later put it, and never came back.

His physics work already had an algorithmic pulse. With Werner Krauth and David B. Wilson, Bernard introduced event-chain Monte Carlo methods for hard-sphere systems. Instead of nudging one simulated particle and frequently rejecting the move, an event chain can displace a sequence of particles in a rejection-free sweep. He also co-authored work on two-dimensional melting that settled the nature of a long-debated transition in the hard-disk model. A later review described the event-chain approach as growing from a clever method for spheres into a broader framework.

After completing his doctorate in 2011, Bernard conducted postdoctoral research at MIT. By 2012, he was imagining an automated machine-learning startup. He did not rush out to make one. Wolfram Research offered problems that were, in his account, too much fun to leave.

“To me, machine learning has always been about being automatic.”Etienne Bernard

Eight years hiding the knobs

At Wolfram, Bernard worked for eight years and led the machine-learning group for seven. One favorite project was Classify, a Wolfram Language function that accepts data and returns a classifier. Its apparent simplicity concealed difficult decisions about algorithms, data preprocessing, hyperparameters, and failure modes. Bernard's memorable analogy was that people do not tune the hyperparameters of a human student, so they should not have to do it for a machine student either.

He also helped develop a high-level neural-network framework. Networks had to be represented symbolically, visualized clearly, and manipulated as objects whose layers could be cut, replaced, or joined. This was engineering as interface design: the point was not merely to expose capability, but to give the capability a shape that a person could reason about.

A 2017 account from Wolfram's summer camp called Bernard the company's machine-learning guru and singled out his 70-minute introduction to artificial intelligence as a favorite lecture. The compliment fits what came next. Before leaving Wolfram, he spent two years writing Introduction to Machine Learning. The finished book uses reproducible examples and deliberately minimizes mathematics. Bernard has said the hard part was deciding what belonged, in what order, and whether an idea was best served by a formula, a piece of code, or a picture.

A company with a cousin and a constraint

Large language models gave Bernard's old startup impulse a sharper opening. GPT-2 and GPT-3 could understand and generate language with a fluency that surprised him. At the same time, building a custom natural-language system remained laborious: annotation teams, expert experimentation, and repeated training cycles. He saw a chance to use large models to simplify the making of smaller, task-specific ones.

Samuel Bernard saw it too. Samuel is Etienne's cousin, and together they founded NuMind. The company entered Y Combinator's Summer 2022 batch with a tool for creating custom NLP models. A private beta launched that August. In March 2023, NuMind announced a $3 million seed round led by Flybridge, with backing from Y Combinator, Big Bets, Carya, Pioneer, Velocity, Sharpstone, and angel investors.

NuMind's public focus gradually narrowed toward documents. The market is vast but gloriously uncinematic. Companies have contracts, forms, invoices, spreadsheets, scans, and reports. Humans can read them, but software needs the meaning converted into a predictable shape. For structured extraction, that shape is usually JSON. For content extraction, often called OCR even when the original file is digital, it may be Markdown.

Small models, exact jobs

Bernard's case for specialization is practical. A general model must be ready to write a poem, debug code, summarize a meeting, and identify a number on a badly scanned form. A document model can spend its capacity on layout, field boundaries, missing information, schemas, and faithful transcription. Smaller size can also make private deployment more manageable, an important point for organizations that cannot send sensitive files to a third-party model provider.

In July 2025, NuMind introduced its NuExtract platform, a web and API product for turning documents into structured output. Users define a template, add examples, test difficult documents, and deploy a task-specific endpoint. The company also offered private installations. The model underneath, NuExtract 2.0 PRO, was designed for multilingual text, PDFs, spreadsheets, and scans.

A template sounds almost quaint beside a language model, yet it gives the system a contract. If the user asks for a customer name, an invoice date, and a list of line items, the output must preserve those fields and their types. Teaching examples then handle the local peculiarities that no generic prompt can anticipate: an unusual label, a recurring page layout, or a code that matters only inside one organization. The glamour belongs to the model; the utility lives in this negotiation between a fixed schema and a variable document.

Bernard's research background also shows in NuMind's habit of publishing comparisons, describing training procedures, and releasing selected models openly. The company has moved through entity recognition, text extraction, visual documents, and reasoning OCR while keeping a public trail of model cards and technical posts. Openness does double duty here. It invites developers to inspect and adapt the work, and it turns adoption into feedback for a small team competing in a field where the general-purpose model makers have much larger budgets.

The May 2026 release of NuExtract3 joined the two branches. The 4-billion-parameter vision-language model handles schema-driven extraction to JSON and full-content extraction to Markdown. It adds optional extraction-specific reasoning, free-form instructions, in-context examples, and a larger vocabulary of field types. Built on Qwen3.5-4B and released under the Apache 2.0 license, it came with public inference and fine-tuning code.

This is where the founder's old preferences become visible in the product. The template makes the job explicit. Examples let users teach by showing. Typed fields reduce cleanup. Private deployment makes architecture part of the value proposition. Returning null when the evidence is absent is a small act of restraint, one that matters more in a database than a clever turn of phrase.

The pedagogy of saying less

Bernard has now spoken at more than 100 events, according to his speaker profile, including SXSW, ODSC Boston, ML Prague, WeAreDevelopers Berlin, and DevTalks Romania. He has guest-lectured at Harvard and MIT. His public explanations tend to begin with the job rather than the architecture: a document goes in; useful, structured information comes out.

There is wit in the simplification. In one Christmas post, Bernard recalled asking Santa for a motorbike and a particular radio-controlled car as a child. The wrong car arrived, and the motorbike did not. Thirty-five years later, while working on an AI startup, he proposed the obvious diagnosis: Santa had a document-information-extraction problem converting children's letters into standardized orders. Somewhere north of the Arctic Circle, a JSON schema had failed him.

The joke lands because it compresses NuMind's pitch without sanding off its strangeness. Civilization runs on documents, and documents remain surprisingly resistant to software. Bernard's career suggests that the way through is neither to celebrate complexity nor pretend it has vanished. Put the complexity in the tool. Give the user a clear surface. Teach the system when to answer, when to abstain, and what shape the truth should take.

The interesting machine is not merely the one that can read. It is the one that knows what the reading is for.