We build multimodal data infrastructure
for frontier models.

Frontier models are moving beyond text. The next generation of systems will need to reason across vision, sound, time, and tools, but that requires a new kind of data infrastructure.

Seldon builds scalable multimodal data and evaluation infrastructure for frontier AI labs. We create human-calibrated tasks, preference data, verifier-style rewards, and agentic judge traces across image, video, audio, text, and tool-use environments.

A core part of our work is long-context video. We turn hours-long source material into evidence-grounded QA, temporal reasoning tasks, and RL-style environments for training and evaluation. Our agentic labeling pipeline combines state-of-the-art automation with human-in-the-loop review, producing some of the largest and highest-quality video datasets in the world. You can read more about our research here.

On top of that, we have built an agentic system that reasons over these massive datapoints and draws coherent, causal conclusions from this multimodal data. We detect micro-trends, track sentiment, and identify target groups at societal scale. See what people think here.