# OnDeck AI - Full reference

> OnDeck AI is a visual intelligence research company building the world's most
> powerful vision model. Our models solve complex video analysis tasks with no
> training and no data labelling - find visual behaviours, structural damage,
> and anomalies in footage in seconds. Founded 2022, based in Vancouver, Canada,
> backed by Y Combinator, SOC 2 Type II compliant.

This document expands on https://ondeckai.com/llms.txt with per-page prose for
answer engines and AI agents.

## Home - the world's most powerful vision model

OnDeck builds intelligence that solves complex video analysis tasks with no
training and no data labelling. Core capabilities:

- Search: find anything across petabytes of unstructured footage.
- Understand: reason about complex events and why they happened.
- Report: generate human- or machine-readable reports for any task.
- Customize: OnDeck self-adapts to each customer's unique workflow.

Benchmark (LVBench overall score): Perception-0 81.2, Gemini 3.1 Pro 78.5,
Deep Video Discovery 74.2, Seed1.5 64.6, AdaReTaKe 53.3, GPT-4o 48.9.
Perception-0 is guaranteed to maintain state-of-the-art over time because it is
a harness over frontier models rather than a fixed trained model.

Enterprise-readiness: deployed with defence agencies, universities, robotics
companies and public safety organizations; runs in the cloud or on-prem with
air-gapped capabilities and SOC 2 Type II compliance; agentic systems
recursively self-improve for each partner's real workflow.

Customers: Singapore's Ministry of Defence (threat detection and situational
awareness across autonomous assets); Parametric (robotics - fine-grained
behaviour understanding to accelerate training of safe robotics models).

## Thesis - why adaptive perception

The problem: visual perception is the most information-dense signal for humans
and machines, yet vision AI has not reached human-level performance even as LLMs
surpass humans in many domains. To understand diverse, dynamic visual data like
humans, vision AI must gain the ability to adapt like the human mind.

The solution: OnDeck builds recursive self-improving mechanisms for vision - (1)
powerful visual analysis primitives and tools, and (2) a system that
autonomously builds a fully formed expert model for any domain. The result is
Perception-0, a harness that gets remarkably better performance out of frontier
models and autonomously self-improves on your task.

The future: as society automates, everything will need visual intelligence.
OnDeck is building a self-improving visual reasoning engine that can solve any
visual task as well as or better than a human, autonomously - a path to being
the last vision research company.

## Enterprise

Extract everything from your footage without training a model. Find visual
behaviours, structural damage, and anomalies in seconds. If a human analyst can
describe it, OnDeck can find it - detection (cracks, corrosion, policy
violations, unusual behaviour, novel objects), reasoning (interactions,
sequences, causes), and memory (understanding across time, with detections
timestamped, clipped and ready to export).

## Defence

The most powerful visual intelligence model for ISR (intelligence, surveillance
and reconnaissance), deployed air-gapped, on-prem and private-cloud for complete
data sovereignty. Trusted by the Singapore Armed Forces. Designed for real-world
operational complexity: unseen targets (identify and describe objects, events
and behaviours beyond a fixed set of trained classes on first encounter), no
training (deploy against new requirements with SOC 2 compliance without
collecting and labelling new datasets), and full temporal behaviour reasoning.

## Environmental

Protecting the planet with 100x faster visual monitoring, trusted by
environmental organizations across the globe.

## Careers

OnDeck is hiring to build the final generation of visual intelligence. See
https://ondeckai.com/careers.

## Research

Peer-reviewed research published at NeurIPS. Paper: https://arxiv.org/abs/2412.02262.
Off-the-shelf VLMs struggle with unseen data; OnDeck's early research enhanced
reliability by grounding predictions in real data rather than model guesses.

## Contact

- Talk to our team: https://cal.com/team/ondeck-ai/intro
- LinkedIn: https://www.linkedin.com/company/ondeckai/
- Twitter/X: https://x.com/OnDeckAI