Skip to content
NB Tech NewsSubscribe

News3 min read

Saint John's Versos Puts NVIDIA Agents on Video Libraries — and Takes Them to Amsterdam

The company launched a plain-language curation layer on NVIDIA NeMo and Nemotron Ultra. It is demoing the stack this week at IBC2026, with Chris Keevill on the Ignite Stage Saturday.

A plain-language search bar, a video library grid with three matching clips highlighted, and a curated dataset of those clips, captioned IBC2026, Amsterdam

Versos AI announced Wednesday a new layer on the Saint John platform: type the video dataset you want in plain language, and agents built with NVIDIA NeMo go find, grade, and assemble matching footage. The company is showing it this week at IBC2026 in Amsterdam — Sept. 11–14 — with CEO Chris Keevill on the Ignite Stage Saturday at 11:45 a.m.

That is a product announcement, not another seed recap. We covered the company in July on the $1.85-million round, the CuriosityStream deal, and the Tape Ark vault-digitization alliance. Wednesday's release is the first time Versos has put a conversational front end on that library-intelligence stack.

Ask for the reel the way you would ask a colleague

Training-data buyers do not want "video." They want a subject, an action, an environment, plus duration, resolution, format, language, production quality, and a licence that will survive a lawyer. Versos's own release says that usually means translating one spec into many searches, staring at results, and packing a delivery by hand.

The new workflow is supposed to collapse that into one conversation. The system turns the prompt into structured criteria, searches the company's scene- and frame-level index, grades candidates against the spec, and builds a dataset. Rights and provenance stay attached; it only searches footage prepared for controlled AI licensing.

"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," Keevill said. "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."

Entrevestor relayed the same quote on Thursday and noted Versos joined the NVIDIA Inception accelerator in June. Inception is a startup program — developer resources, training, preferred pricing — not an investment announcement.

Three layers, one model they can swap

The company names the stack in public:

  • NVIDIA CUDA Toolkit accelerates the inference that turns raw footage into scene- and frame-level data inside the Video Library Intelligence Platform.
  • NVIDIA Nemotron Ultra runs the agents that interpret the request and do the discovery and curation.
  • LangChain coordinates the steps between the prompt and search, metadata, grading, and assembly.

The architecture is model-agnostic on purpose. Nemotron Ultra is what runs today; the same workflow can sit on other open-weight or frontier models if a customer wants a different cost, performance, or deployment story. That is the part of the release that is not a logo slide: they will take NVIDIA's agent model, and they will not weld the product to it.

Amsterdam is the showroom this week

IBC is the broadcast and media-tech trade show. Versos is in the Future Tech Ignite Zone, stand 14.FT37, Hall 14. Keevill's Saturday talk is the public hour. The company is taking meetings off versos.ai/contact.

None of that is a purchase order. It is a Saint John company putting a new interface on licensed video data in the room where studios and AI buyers actually walk. The July piece was the financing and the partnerships. This week is whether a plain-language prompt can pull a rights-clean reel off a library without a week of human search.

Sources

Related coverage

Tags