One foundation. Many forms.
A connected set of intelligent products that share a spine — so intelligence follows the person instead of being trapped in an app.
A hundred boxes that do not speak.
The software industry settled on a shape: one app, one problem, one silo. It worked when software only followed instructions. It breaks the moment software can reason, because reasoning needs context — and context is exactly what a silo destroys.
The result is familiar. Every tool knows one narrow thing about you. None of them knows you. You spend your day being the integration layer between systems that should have been talking to each other.
Six layers, one spine.
Surfaces
software and devices
The things people actually touch — some pure software, some software living inside hardware. Intelligence is worthless until it meets a moment. Surfaces are where the meeting happens.
Orchestration
routing, tools, task state
Routing between models, breaking a request into steps, calling the right tools, holding state across a long job, knowing when to escalate. This is where most of the real work happens. A single model answering a single question is not a product.
Intelligence
models, reasoning, memory
The models, the reasoning, and the shared memory every surface draws on — so a product does not start from nothing every time, and so what is understood in one place is useful in the next.
Trust
identity, policy, audit
Identity, permissions, encryption, data residency, auditability — built into the layer, not bolted onto each product. A system that knows this much about a person or a business has to be answerable for it.
Foundation
inference, on-device runtime, silicon
The engineering underneath — inference, on-device runtimes, quantisation, the silicon integration work. The unglamorous part that decides whether any of the above is actually possible.
Evaluation
quality, drift, harm
Continuous measurement of quality, drift and harm, feeding corrections back into orchestration. The only layer that points back upward. It is how we know the system is still right, rather than assuming it.
We build down as far as the problem requires.
Most of what we make is software, because most problems live where the work lives. But some do not. When a problem needs a response faster than a network round trip, or data that should never leave a room, or a sensor that no phone has, software alone cannot answer it. So we build the hardware and put the intelligence inside it.
This is a deliberate constraint on ourselves, not a product line. We do not build hardware because hardware is interesting. We build it when the alternative is a compromise the person would feel.
| Pure software | Software integrated with hardware | |
|---|---|---|
| When | The problem lives in systems that already exist | The problem lives in a physical place or moment |
| What it gives | Reach, speed of iteration, no new object in anyone’s life | Latency measured in milliseconds, data that never leaves, capability no general-purpose device has |
| The bar | Must be useful the first time it is opened | Must be worth the space it occupies |
The part nobody sees.
Anyone can ship a list of products. An ecosystem is the harder claim, and it comes down to three things being genuinely shared rather than separately rebuilt: how the system reasons, what it is allowed to remember, and who is allowed to see it. We build those once, properly, underneath everything.
The test is simple. If you could remove a product from the set and nothing else got worse, it was never an ecosystem.
Complexity is our side of the line.
Everything described above — six layers, shared memory, on-device inference, governed permissions — is machinery. None of it should ever surface as something a person has to configure, understand or maintain. An ecosystem that makes its architecture the user’s problem is just a platform with extra steps.
The measure we hold: the more capable the system becomes, the less it should ask of the person using it. If those two lines ever move in the same direction, something has gone wrong in the design and not in the engineering.