Machines and devices
Runs on the device itself: transcription, reading images, sorting.
Quantised local models
NotA cloud call from a device that keeps losing signal.
AI and agents
AI, in the places it actually pays.
Anyone can buy the same model you can. The work is connecting it to your systems, setting what it may do, and showing where an answer came from.
- Shipped
- Platform · connectors · agents
- Hosting
- Cloud · on-prem · air-gapped
- Foundation
- Our own open-source stack
What we build
Eight kinds of build, one foundation.
01
Intelligent platform
One place your team works: projects, channels, saved sessions.
02
An assistant
Answers from your systems, not the open web.
03
Searchable documents
Your documents, searchable by meaning rather than keyword.
04
Connectors to what you run
The systems you already run, opened up for an assistant to query.
05
Automation and workflow
Long processes that survive a crash or a restart.
06
Analysis and forecast
Charts and forecasts from live data, not last month’s export.
07
Harness engineering
The rules your developers’ AI has to work inside.
08
Training your team
Workshops, from how a model works to running agents safely.
In production
No internet, no documentation, one engineer.
An industrial group whose machines cannot reach the internet, and no record of what runs where.
We built a searchable, read-only picture of everything they run, on models hosted on their own machines. Nothing leaves the site.
A technician pastes a screenshot of an error. The answer comes back with a link to each thing it checked.
Placement
Where it belongs, and where it does not.
The software you already run
An assistant that queries your systems and shows what it read.
MCP servers over existing systems
NotPasting a database layout into a prompt and hoping.
Moving data between systems
Proposing how two systems’ fields line up, and where they clash.
Structured output · typed contract
NotLetting it write into your ERP unsupervised.
Services and APIs
What the model returns is checked at the door, like any outside input.
JSON Schema · OpenAPI 3.1
NotTrusting the model’s text because it looked right.
Reporting and forecasts
Forecasts, unusual patterns and charts from your own data.
Query shown · result reproducible
NotA number nobody can trace back to a query.
Your development team
Agents inside set limits, every change checked before it lands.
Agent Skills · fitness gates
NotAutocomplete pointed at production and called a strategy.
Keeping it running
Working out what broke, from a map of what you run.
Per-call audit · citations · SSO
NotAn unaudited assistant holding production passwords.
The foundation
We already own the layer underneath.
Most of what an AI product costs is plumbing every project rebuilds. Ours is already built.
- L1 · Runtime
- Nothing reaches a file or the network without an explicit grant.secure by default · no node_modules · single binary
- L2 · Framework
- Client, server, validation and documentation derive from one definition, so an integration error is a compile error.typed contracts · plugins · AI stack · traces
- L3 · Product
- The part that is specific to your business.your domain · and nothing else
We publish that middle layer as NetScript, openly licensed.
Harness · the run loop
Five steps, and one way back.
Work that fails a check comes back.
Read the rules before starting
The agent reads the architecture rules and the known problems first.
Explain the plan before writing code
Before any file exists, the agent states what it is going to build.
Build it in the pieces it named
Going off that plan is reported rather than allowed through quietly.
Eighteen automatic checks
They pass or they fail. There is nothing to argue with.
Reviewed by a second agent
The reviewer starts fresh, with no memory of the work, and returns a verdict.
Governance
What a security review actually asks.
Pilots stall at the same six questions every time.
- Model access
- One gateway, keys held server-side, bring-your-own-key per team.
- Data boundary
- One component talks to the model provider. Nothing else does.
- Identity
- Single sign-on over OIDC and existing group membership.
- Audit
- Every conversation kept as the record, including what was read.
- Cost
- Tokens metered per channel, provider and model.
- Sovereignty
- Swiss or EU hosting, on-premises, or fully air-gapped.
Next step
Start with the problem, not the model.
Tell us which job you want AI to do, and which systems it would read.