Versos Dataset-Building Agents for AI Model Training: Built on top of NVIDIA Nemotron 3.5 Super VL

When an AI company sends you an RFP, the request is specific and the timeline is short. Finding what matches in your library, assembling it into a dataset, and getting it to the buyer is where these deals are won and lost.

The Versos Library Intelligence Platform is the infrastructure media companies use to prepare their data, build the dataset AI labs are asking for, and deliver it. To make things easier, Versos has built an agent on LangChain that behind the scenes uses NVIDIA Nemotron 3.5 Super VL, a Vision-Language Model (VLM), to guide users through the entire workflow. What used to take a team of engineers and data scientists is now made simple and fast.

Building bespoke datasets to highly refined specifications is hard to do at scale and at speed. With the Versos agent, it is minutes away. Tell your agent what you want to deliver, and the agent builds and delivers it for you.

How It Works

1. Start with the request spec:
   Paste in the spec or describe the requirements in plain language. Your Versos agent extracts the criteria that matter, including subjects, actions, environments, duration, resolution, language and permitted use.
   
   
2. Let the agent find what you can't:
   Your Versos agent translates your request into its own search strategy, generates the queries it needs, and explores your content at scene- and frame-level detail. It can find content you may never uncover, and builds the view needed to make sense of it. In seconds, you see what your library can actually deliver.
   
   
3. Showcase your data:
   Versos assembles the content that meets the specification and builds a gallery review interface you can share with the buyer. See the titles, scenes, clips and metadata behind every selection before anything leaves your library.
   
   
4. Deliver it:
   The dataset is delivered to the buyer's cloud, carrying the metadata the buyer's training pipeline requires. What used to take weeks of coordination becomes a conversation with an agent.

Powered by NVIDIA Nemotron 3.5 Super VL

Versos built the agentic curation workflow on NVIDIA technology, from the agents to the chips:

●      LangChain provides the agent scaffolding layer that coordinates the steps between a plain-language request and the platform's search, metadata, grading, and dataset-assembly functions. The approach supports multi-step workflows, tool use, and specialized-agent delegation while preserving the modularity needed to integrate models and infrastructure based on customer requirements.

●  NVIDIA Nemotron 3.5 Super VL powers the agents that interpret a buyer's request and enables them to carry out discovery and curation of the video training data. The Nemotron 3.5 Super VL is an open multimodal model for long-running agents that need to see, reason, and act.

●  Nemotron 3.5 Super VL model combines text, image, document and video understanding with high-accuracy agentic reasoning and tool use in one model. The agent keeps task-relevant context (with 1M context window) from the buyer's spec through search, review and final selection, and its native video understanding extracts the key information from footage that curation depends on.

●  Because Super VL is open, Versos can customize it with NVIDIA NeMo framework using its own data, tools and workflows, and deploy it across local, data center and cloud environments while keeping control of model behavior and data handling.

●      The NVIDIA CUDA Toolkit accelerates the inference that runs inside the Versos Library Intelligence Platform, the analysis that turns raw footage into scene- and frame-level data.

●      Versos runs on NVIDIA GPUs across four generations: B300, B200, H200, H100, A100, L40S and L4.

●      As Versos expands the platform, it plans to evaluate NVIDIA's newer agent-execution and observability capabilities, including NeMo Fabric and NeMo Relay, to add consistent harness execution, normalized workflow results, traceability, middleware, guardrails and model-routing controls. Versos also expects to assess relevant capabilities from the broader NVIDIA Agent Toolkit ecosystem as they become applicable to video-data curation.

Get in touch to see how Versos uses NVIDIA-powered AI agents to help media companies build and deliver custom video training datasets in minutes, not weeks.