Pragma Research / Investors
Cadence is a novel real-time AI modeled after animal and human brains. It learns from live experience. Our ambition is to put it inside robots that keep improving throughout their working lives.
Proposed seed round: $4 million over 18 months.
Working software and live demonstrations. Seeking our first robotics evaluation partner.
The gap
Robots work in a world nobody trained them for. A model that is frozen at the factory cannot learn from any of it.
Transformers
Most of today's AI runs on transformers. A transformer is completely feed-forward, and in a typical deployed model its attention matrices and all other weights are fixed once training ends. New knowledge means fine-tuning or retraining, a separate run in a data centre.
Cadence
Cadence follows the design of a brain. It senses, decides and acts in real time, and each action and its result changes it. The brain that works is the brain that learns.
Cadence is built from small self-reading patches. Each one holds its own state, listens to its neighbours, keeps records of experience and corrects itself through feedback. Together they settle on one answer, and experience changes how they are connected. For a robot this means learning on the job, adapting when its body changes and acting only on a settled answer.
How Cadence works →What exists today
Software
Cadence is working software. Its core is open source, and anyone can install it and build a brain.
Cadence source →Demonstrations
Brains start untrained in a browser tab and learn while you watch. They play Atari games, evolve in a simulated world and recover when a rover's wheel is weakened.
Watch it learn →Next
Everything so far runs in simulation. The seed round takes the brain onto a real machine, with a robotics partner and a measurable task.
The proposed round ↓The opportunity
Robotics is our first market. The proposed product is a supported Cadence runtime and a brain prepared for the customer's embodiment, with a path from one useful integration to recurring fleet revenue.
A changed payload, a worn motor or a different work surface can turn a familiar action into a failed one. Our first customer test is whether a learning brain reduces the operator interventions that follow such a change.
01
Integrate one valuable task with a robot maker. Measure how the robot adapts against the maker's existing stack.
02
Convert useful results into a runtime and embodiment license, with deployment integration and support.
03
Expand through additional robots and tasks using the same brain architecture and reusable integration tools.
The Cadence core is available under GPL-3.0. Our proposed commercial value is a supported runtime, prepared brains, integration tools and the expertise to make learning useful on a customer's hardware. Commercial licensing arrangements will be established during product development.
Proposed seed round
We are seeking a lead investor and participants to turn the working software into a deployable robot brain. The proposed programme targets three outcomes.
01
Show a real robot adapting to a change that defeats a frozen brain. Measure how fast it recovers, what it retains and what it costs to run.
02
Complete a paid evaluation where learning reduces intervention or increases productive operation.
03
Transfer the software to a second task or integration and establish a path to a deployment license.
Staffing, hardware access and milestone budgets form part of investment diligence.
Founder
Creator of Mythril, the symbolic execution tool for Ethereum, and a creator and lead author of the OWASP mobile security standards and testing guide.
Bernhard leads Pragma Research and the development of Cadence. The proposed round adds robotics integration and systems engineering capacity alongside learning research.
Inspect the work
The open-source library with documentation and runnable examples.
Cadence source →Every demo on this site with its full source and recorded measurements.
Demo source and results →The Cadence and physics preprints, with links to the underlying mathematical work.
Preprints and research →Performance on a physical robot, deployment economics and customer demand are the next tests.