Neuralk-AI

Here is a fact that AI enthusiasts find slightly annoying, which is usually a sign that it is important: the most valuable data inside most companies is boring. It is not a clever essay or a photorealistic image. It is a table. Rows and columns. A product catalog, a list of customers, a spreadsheet of transactions that some analyst cleaned up in 2019 and has been quietly running the business ever since. Large language models, for all their charisma, are strangely bad at this stuff. Ask a chatbot to reason over a 400,000-row SKU file and it will do something confident and wrong.

Neuralk-AI, a Paris deeptech founded in 2023 by Antoine Moissenot and Alexandre Pasquiou, decided to build for the boring thing on purpose. Its pitch is that tabular data - the structured backbone of commerce - deserves its own foundation model, the same way text got GPT and images got diffusion. And the reason this is a real business rather than an academic footnote is that the tables in question are the ones deciding what a retailer buys, how it prices, and what it recommends. When those decisions improve by a few percent, the numbers get large.

Pasquiou, the chief scientific officer, has a nice way of framing why the opportunity sat there unclaimed. "The technology underpinning tabular data has barely evolved in decades," he says, "particularly in retail where structured datasets drive every key decision." This is the kind of sentence that is easy to nod along to and then forget, so it is worth sitting with. The tools most data teams reach for - gradient-boosted trees like XGBoost - are genuinely good and genuinely old. Everyone assumed the problem was solved. Neuralk-AI's entire thesis is that "good enough and old" is exactly where you find a startup-sized opening.

The science underneath has an unusual origin. Pasquiou holds a PhD in computational neuroscience - the study of how brains represent information - which turns out to be a weirdly apt background for a company obsessed with how machines should represent a table. Moissenot, the CEO, is an "X," which is French shorthand for a graduate of École Polytechnique, the country's most prestigious engineering school. The combination reads like a French deeptech central-casting call, and the investors clearly agreed.

Because the other thing that happened, in February 2025, is that a lot of people who know AI wrote checks. Neuralk-AI raised roughly $4 million in a seed round led by Fly Ventures, the Berlin firm that backs early enterprise and deeptech founders. The angel list is the tell: Thomas Wolf, co-founder of Hugging Face, who said the company could "unlock the untapped potential of tabular data"; Charles Gorintin, the co-founder of the health insurer Alan; and Philippe Corrot and Nagi Letaifa of Mirakl, the marketplace software company. When the people who built the reference platforms for open AI and for online marketplaces both put money into your tabular startup, you have at least chosen an interesting problem.

Neuralk-AI aims at companies that already have data warehouses and data science teams - which is to say, enterprises with a mess big enough to be worth cleaning. The early testing partners are a who's-who of French commerce: the retail groups E.Leclerc and Auchan, the marketplace platform Mirakl, and the retail-media startup Lucky Cart. Mirakl, whose executives are also investors, has been trialing the model on a deeply unglamorous but expensive problem: eliminating duplicate entries in product catalogs.

The company says it has mapped around 45 distinct use cases across commerce - deduplication, catalog enrichment, assortment optimization, demand forecasting, pricing, personalization, fraud detection. None of them make a flashy demo. All of them, summed across a large retailer, are the difference between a good quarter and a bad one. That is the unfashionable, durable kind of value proposition, and it is a deliberate choice.