Versos AI announced Wednesday it has developed an agentic system using NVIDIA technology to help artificial intelligence companies search for and assemble video datasets for model training.

The Saint John company is demonstrating the new capability at IBC2026 in Amsterdam this week. The system allows users to describe the type of video footage they need in plain language, after which the platform translates the request into structured criteria and searches video at the scene and frame level.

The system evaluates potential footage against requirements such as subject matter, actions, environments, duration, resolution, format, language, production quality and permitted uses. It then assembles footage that meets the specified criteria into a structured dataset.

“AI teams should be able to search for the data they need to train new models the same way they would describe it to a colleague,” said Versos Co-Founder and CEO Chris Keevill. “The people building AI deserve to benefit from it just like everyone else, and our agentic interface makes that happen by streamlining the process for curating training data.”

Versos, which joined the NVIDIA Inception accelerator in June, said the system is designed to address a process that can involve manually searching large video libraries, reviewing individual clips and organizing selected footage for delivery.

The workflow also incorporates licensing and provenance information. It searches footage that has been prepared for controlled AI licensing while maintaining information about ownership, licensing and the source of the content.

The agentic system was built using NVIDIA and LangChain technologies. NVIDIA's CUDA Toolkit is used to accelerate inference within Versos' Video Library Intelligence Platform, which analyzes footage at the scene and frame levels.

Versos said its architecture is model-agnostic. Although Nemotron Ultra is being used to run the agents, the workflow can also operate with other open-weight or frontier AI models, depending on customer requirements for performance, cost or deployment.

For video library owners, the system is intended to provide a way to respond to specific requests from AI companies without manually reviewing an entire library for every request.

Versos AI, which closed a $1.85 million equity round last December, was founded to provide infrastructure for licensing video content for AI model training. Its Video Library Intelligence Platform analyzes and indexes video at the frame level, allowing content to be searched and organized for potential AI applications.

The company will demonstrate the technology at IBC2026 from Sept. 11 to 14. Keevill is scheduled to speak at the conference's Ignite Stage on Sept. 12.