EGM AI
We build the parts of an AI system that a demonstration leaves out: the deployment, the retrieval over your own documents, the controls on what a model may do, the audit trail, and the integration into the software your people already use. We have run this infrastructure ourselves for years, on our own hardware, before offering it to anyone else.
Local
Models running on your own hardware, inside your own network, with nothing leaving the building. For organisations whose data cannot be sent to a provider: regulated sectors, schools, legal and financial work, anyone holding records that a contract or a regulator says must stay on site.
We specify and commission the hardware, build the inference layer, tune it for the models you will actually run, and hand over an operator guide your own IT function can follow. No telemetry, no subscription, no dependency on a vendor's roadmap.
Cloud
Frontier models from the major providers, used under governance rather than by default. For organisations whose data can be processed externally under contract, and who want the strongest available models without giving up control of what goes to them and what comes back.
We design the boundary: what may be sent, what must be stripped first, what a model may act on and what it may only recommend. Then we build the system around it, with the same audit trail and the same human oversight as a local deployment.
System design and readiness
Which of your workflows an AI system can carry, which it cannot yet, and what the deployment would cost to run. A written decision record, including the answer do not build.
Local inference deployment
Hardware specification, model selection, the serving layer and its tuning, and a tested operating procedure. Multi-machine where the load requires it.
Retrieval over your documents
Systems that answer questions from your own records with the source shown, tested adversarially before they are trusted, and reported on for what they can and cannot be relied on to do.
Agents with controls
Models that act inside your systems through a gateway that enforces what they may touch, at what permission tier, with a kill switch and a complete record of every action taken.
Integration
AI capability placed inside the software your people already use, rather than a separate tool they must remember to open.
Cloud AI under governance
Provider selection, the data boundary, prompt and output handling, cost control and audit, for organisations using external models under contract.
Private assistant systems
A local-first assistant for a person, a household office or a small company, with retrieval, memory and human oversight, running across your own machines.
Operator training
The people who will run the system are trained to run it, through our Education and Training division, so that it does not depend on us after handover.
The company's own assistant system has run continuously on a cluster of four machines for over a year. It is the environment in which our methods were developed and it is the reference for what we build for clients.
It is also where our finance platform is being built: a system for managing a business's finances entirely through local models, for owners who will not put their ledger through a cloud provider. Both are in daily use and under continuous development.
Orchestration. One machine directs the work; three others carry inference, retrieval and utility tasks, joined over a private network.
Inference. Models of up to 120 billion parameters served on the company's own hardware, selected and tuned for each task rather than one model for everything.
Retrieval and memory. A private document store with dense retrieval, and a memory plane the operator controls, so that the system knows what it has been told and nothing it has not.
Controls. A tool gateway that decides what the model may act on, permission tiers that separate reading from doing, and an autonomy layer that stays off until the operator authorises it in writing.
Depth. Beneath this, the company is building its own inference engine, an execution layer for large language models drawing on published research and the strongest open codebases, so that our understanding of the stack goes to the bottom of it.
- Position uncontrolled automation as a strategy. Every system we build has a person in the loop where a decision lands on someone.
- Send your data to a provider without a written boundary that says what goes, what is stripped first, and why.
- Sell you a model. We sell a system, and the model inside it changes when a better one exists.
- Build something that only we can operate. Every deployment ends with an operator guide and a trained operator.
What the work is, who does it now, what records it touches and where those records are allowed to go. From that we can say whether a local or cloud deployment fits, and what a first engagement would cost.
Write to the company