Tabular Data and Artificial Intelligence

Introduction to Tabular Data

Tabular data is a broad term that encompasses structured data that generally fits into a specific row and column. It can be a SQL database, a spreadsheet, a .CSV file, etc.

Limitations of Large Language Models

While there has been tremendous progress on artificial intelligence applied to unstructured and sequential data, large language models (LLMs) are fuzzy by design. They are built to manipulate input tokens to generate a coherent output without necessarily following a fixed structure. The best LLMs are also either expensive to access via an API or expensive to run on your own cloud infrastructure.

Existing Data Strategies

And yet, many companies already have a data strategy with a data warehouse or data lake to centralize all important data and some data scientists that can leverage this data to improve the company’s strategy.

Neuralk-AI and Tabular Data

French startup Neuralk-AI is an artificial intelligence company that has been working on AI models focused on tabular data. The company recently announced $4 million in funding.

Focus on Structured Data

“Data with real value for companies is data that was identified a long time ago, structured in the form of a table, and used by the data scientists of these companies to create all their machine learning algorithms,” Neuralk-AI co-founder and Chief Scientist Officer Alexandre Pasquiou told TechCrunch.

Opportunity in Tabular Data

Neuralk-AI thinks there’s an opportunity in revisiting AI model development, but with a specific focus on structured data. At first, it plans to offer its model as an API to data scientists working for commerce companies because these companies love data — think product catalogs, customer databases, shopping cart trends, etc.

Limitations of LLMs

“Today, LLMs are great for search, natural user interaction, and answering questions based on unstructured documents. But it has some limitations the moment we go back to classic machine learning, which is really based on classic tabular data,” Pasquiou said.

Applications of Neuralk-AI

With Neuralk-AI, retailers can automate complex data workflows with smart deduplication and enrichment. But they could also use the company’s models to detect fraud, optimize the product recommendations, and generate sales forecasts that could be used for inventory management and product pricing.

Funding and Partnerships

Fly Ventures led the company’s $4 million round with SteamAI also participating. Several business angels also invested in the startup, such as Thomas Wolf from Hugging Face, Charles Gorintin from Alan, and Philippe Corrot and Nagi Letaifa from Mirakl.

Testing and Development

The team is still actively working on its models. It plans to test with a group of leading French retailers and commerce startups, such as E.Leclerc, Auchan, Mirakl, and Lucky Cart.

Future Plans

“Within three or four months, we’ll release the first version of our model and the public benchmark on which we’ll be able to rank our model compared to the state-of-the-art in this space,” Pasquiou said. “And in September, the idea is to be the best tabular foundation model in everything related to representation learning.”