Neuralk launches Seldon to predict structured business data

Neuralk has introduced Seldon, a specialized tabular foundation model designed to handle predictive forecasting directly from structured business data. While large language models like ChatGPT and Claude excel at summarizing spreadsheets, Neuralk CEO Alexandre Pasquiou argues that they obscure the underlying structure of numeric data and struggle to learn numeric distributions. Seldon aims to bridge this gap by learning directly from rows, columns, and numbers, offering a single predictive brain for enterprise workloads.

The Seldon model is built to process massive datasets, scaling to roughly 20 million rows and more than 600 columns, which far exceeds the typical context window limitations of standard language models. Instead of requiring companies to build separate, custom machine learning pipelines for every business question, Seldon acts as a general-purpose engine. It can adapt to diverse analytical tasks such as customer churn, fraud detection, demand forecasting, pricing, classification, and regression.

For practitioners, Seldon integrates directly into existing workflows. It connects through Python, Excel, Model Context Protocol servers, and custom skills, allowing users to hand off predictive tasks from interfaces like Claude or ChatGPT. This setup enables a multimodel approach where a standard language model acts as the conversational interface while Seldon runs the heavy predictive calculations in the background.

Neuralk is offering Seldon for free to start, with hosted on-premises and private cloud deployment options available for enterprise customers with sensitive data. Pasquiou predicts that within the next few years, tabular foundation models will power nearly all predictive workloads in the enterprise, while traditional AI agents focus on automating workflows.