Most tasks don't need frontier intelligence. They need something that knows the job, runs where the work happens, and won't break when a lab ships its next update. We factor intelligence into modules that run on hardware you own.
coprime factors intelligence into two modules: a belief network and a reasoner. The belief network is a faithful, self-revising model of the world. The reasoner has no knowledge encoded in its weights and reasons only over the belief network. This preserves performance and is an order of magnitude lighter to run. The work stays where you are, under your control, and keeps running whether or not the cloud is reachable.
A drone clearing a building, a robot on a line, a vehicle reading a road — these can't always pause to round-trip a data centre. The decision has to happen on the device, in milliseconds, with no connection to depend on.
That's the architecture we are building: light enough to run on the device itself. The belief network keeps a running, revisable model of the world the machine is acting in; the reasoning module thinks over that model. Perception and action later plug into the same two modules. Autonomy runs on-board, not as a fallback for when the cloud is missing, but because that is where it has to be.
This is capability the cloud can't deliver, not a cheaper version of it.
A graph-backed, temporally-aware belief network: what's true, when it was true, and how it's known. It revises itself instead of bloating. No more AI that forgets or contradicts itself.
No knowledge baked into its weights, it decides what to do by reasoning over the world model in latent space. An order of magnitude lighter, and fast enough to run on a phone, drone or vehicle.
Two modules, not one fused model. More reliable, and yours. Every belief is checked against rules it can't break. Nothing leaves; nothing is taken on trust.
Powering SYMPHONIA: the LLM-assisted expert consensus platform to support UK government policy making. In collaboration with ARC Yorkshire & Humber.
Co-building autonomous multi-agent systems for data office operations that reason where the data lives.
Building governed, sandboxed agent stacks: multi-agent systems that run with guardrails, governance and isolation by default.
The next wave of AI will not live in a data centre. It will live in your pocket, at your clinic, on your machines. It gets there by being modular and able to run on the hardware you already own. The belief network and reasoner are where we start; they are built to be pluggable. The first modules in a platform for designing and building edge ML systems that run anywhere.
If it can run on the device, it should. The cloud is the exception, not the default.
A memory that revises itself beats one that only grows. What we know should stay right as the world changes.
Every belief carries its evidence and meets rules it cannot break. Correctness and privacy as architecture, not policy.