Question 1

What is OnDeck AI?

Accepted Answer

OnDeck AI is a visual intelligence research company based in Vancouver, Canada, building the world's most powerful vision model. Its models solve complex video analysis tasks - finding visual behaviours, structural damage, and anomalies in footage in seconds - with no training and no data labelling. OnDeck was founded in 2022 and is backed by Y Combinator.

Question 2

What is Perception-0?

Accepted Answer

Perception-0 is OnDeck's flagship vision model. Rather than a traditional trained model, it is a self-improving harness that extracts substantially better performance from frontier models and autonomously improves on each task. As the underlying frontier models improve, Perception-0 keeps getting better. It scores 81.2 overall on the LVBench long-video benchmark.

Question 3

How does OnDeck work with no training or data labelling?

Accepted Answer

OnDeck combines powerful visual analysis primitives with an agentic system that autonomously assembles a fully formed expert pipeline for any domain. This means you can deploy against new requirements without collecting or labelling new datasets - if a human analyst can describe it, OnDeck can find it.

Question 4

Where can OnDeck be deployed?

Accepted Answer

OnDeck can run in the cloud, on-premises, air-gapped, and in private-cloud environments for complete data sovereignty. It is SOC 2 Type II compliant, making it suitable for defence and other security-sensitive deployments.

Question 5

Which industries use OnDeck?

Accepted Answer

OnDeck is deployed across defence (ISR), enterprise, robotics, universities, and public-safety and environmental organizations. Singapore's Ministry of Defence uses OnDeck for threat detection and situational awareness across autonomous assets, and Parametric uses it to accelerate the training of safe robotics models.

Question 6

How does OnDeck benchmark against other models?

Accepted Answer

On LVBench, OnDeck's Perception-0 scores 81.2 overall, outperforming Gemini 3.1 Pro (78.5), Deep Video Discovery (74.2), Seed1.5 (64.6), AdaReTaKe (53.3), and GPT-4o (48.9). Because Perception-0 is a harness over frontier models, its performance is guaranteed to improve as the underlying models improve.

Question 7

Has OnDeck published research?

Accepted Answer

Yes. OnDeck has published peer-reviewed research at NeurIPS, including work on improving vision-language-model reliability by grounding predictions in real data rather than model guesses (arXiv:2412.02262).

Intelligence that solves complex visual analysis tasks for you.

AI that sees and adapts as well as a human

Search

Understand

Report

Customize

OnDeck outperforms all other models.

We've published with NeurIPS.

Proven. Flexible. Built for enterprise.

Proven

Deployable

Customizable

The most advanced teams use OnDeck.

Singapore's Ministry of Defence improved threat detection and situational awareness across autonomous assets using OnDeck.

Parametric leveraged OnDeck's fine-grained behaviour understanding to accelerate the training of safe robotics models.