Most of the venture dollars flooding into artificial intelligence chase the flashy stuff—chatbots that sound human, image generators that mimic Picasso, agents that promise to replace junior analysts. Neuralk AI wants the tables instead. The spreadsheets. The CSVs that actually run businesses.
The Paris-based startup announced a $4 million pre-seed round in early February, money it plans to spend building what it calls "frontier AI models" for structured data. Berlin's Fly Ventures led the round, with backing from StemAI and a collection of angels that reads like a who's-who of European tech: Thomas Wolf from Hugging Face, Charles Gorintin of the health insurance unicorn Alan, and several executives from marketplace software maker Mirakl. French press accounts suggested that a portion of the funding arrived as debt, though the company hasn't disclosed specifics.
The pitch is straightforward, if somewhat contrarian. While much of the AI world has spent the past two years fine-tuning large language models for every conceivable use case, Neuralk's founders believe enterprises need something built differently for the rows and columns that power their operations. Their API promises instant predictions from business tables—classification tasks, regression analysis, anything that involves numeric fields, text fragments, and categorical data across datasets with up to a million rows or so. Time-series forecasting, they say, is on the roadmap.
When XGBoost Isn't Enough
CEO Antoine Moissenot and Chief Science Officer Alexandre Pasquiou—both products of France's elite technical universities, École Polytechnique and CentraleSupélec—argue their tabular foundation models already outpace workhorse tools like XGBoost and CatBoost. More boldly, they claim to be "twice as effective" as general-purpose LLMs when those models get pointed at structured data problems. The secret, according to the company, lies in pretraining on millions of synthetic datasets, an approach that mirrors how image and language models learn broad patterns before specializing.
Whether that performance edge holds up under rigorous third-party testing remains to be seen. The company has published some benchmarks, but independent validation is still sparse. Then again, so is the competition at this stage.
Neuralk has pulled together early pilots with French retail giants E.Leclerc and Auchan, alongside Mirakl (whose CTO, Nagi Letaifa, doubled as an investor) and coupon technology firm Lucky Cart. Testimonials on the startup's website feature supportive quotes from Letaifa and Mehdi Ghissassi, a director at Google DeepMind, though those endorsements don't yet translate into disclosed revenue figures or detailed case studies.
The company was tapped for the French Tech 2030 accelerator program last November and recently picked up a "Jury's Favorite" nod at the LSA IA Tech for Business Awards in January. Fourteen employees now operate out of Station F, the sprawling startup campus in Paris, with open roles spanning AI research, engineering, and sales.
A Category Taking Shape
Neuralk is hardly alone in sensing opportunity where data lives in neat rectangles rather than flowing prose. The space has gotten noticeably warmer over the past year.
In the U.S., a startup called Fundamental pulled in $255 million last fall at a $1.2 billion valuation, positioning its petabyte-scale forecasting models as purpose-built for tabular data. Germany's Prior Labs has been pushing open-source alternatives with an eye toward commercial deployment. Even SAP jumped in this past January, unveiling a Relational Pretrained Transformer designed specifically for business tables—a sign that incumbents see the writing on the wall, or at least on the spreadsheet.
The question isn't whether structured data deserves specialized models. Most enterprises would agree their most valuable information doesn't live in PDFs or Slack messages; it lives in databases, inventory systems, and financial dashboards. The question is whether "tabular foundation models" become a genuine product category or just a feature that gets absorbed into broader analytics platforms.
For now, Neuralk has enough runway to make its case. The $4 million buys time to refine the models, expand the customer roster beyond French retail, and prove that businesses will swap out their XGBoost pipelines for API calls. If the bet pays off, the startup could carve out a profitable niche in an AI landscape that too often ignores the unglamorous work of parsing enterprise data.
And if it doesn't? Well, there's always the option of pivoting to chatbots. Those still get better press.
