What we do

Private AI, placed where it belongs.

Hardware you control comes first — air-gapped and on-premise. Sovereign cloud takes the work that doesn't justify its own machine. Approved cloud comes last, and only where you allow it. We work out which workload runs where, put in the infrastructure, and build the agents that do the work.

Request an inference audit →See the approach

Inference audit

Work out where AI belongs before anything is built.

We sit down with the people who run the workflow and map it: which systems hold the job files and the rates book, who signs off and on what. What does it cost when the work goes wrong? That gets priced too. The output is a plan you can act on — where AI should run, how it will be judged, and a keep-or-stop gate set up front.

One narrow slice of the workflow selected and scoped
Source systems, decision rights and failure costs mapped
A measurable baseline agreed before the build starts

Agent development and deployment

Build the agent into the workflow itself.

The agents we build do work you already do — price from the tender pack, draft the safety response, keep the register current. They pull from your approved records and act inside the permissions you set. When a decision belongs to a person, they stop and hand it over. What happened stays on record. All of it built against your systems, not a demonstration dataset.

Grounded retrieval from your approved records
Role-scoped permissions and explicit approval gates
Exception handling and a reviewable audit trail

Operational knowledge

Connect AI to your records without losing control of them.

Your procedures, contracts and job histories already live somewhere — SharePoint, network drives, the ERP. We connect AI to them where they sit. Every answer carries its source, so the person checking it can open the original. Access runs on the identities and permissions you already administer, and each deployment is designed to respect who may see what, which revision is current, and what each source may be used for.

Running today: sources stay attached to every answer
Designed in: permission-, revision- and classification-aware retrieval
Delivered inside the standard engagement — no separate knowledge-platform rollout

Governance & provenance

Control you can show a reviewer.

Somebody will eventually ask. Which document was that priced from? Who approved it? What happened next? Every workflow we deliver ships with permissions, approval gates and an audit record, so the answers are already there. Provenance shows where an answer came from — your people still decide whether it's right.

Running today: the five controls in every delivered workflow
Designed in: records a reviewer can walk from output back to source
In development: deeper lineage across models, agents and tools

AI infrastructure

Deploy the compute the work requires.

The machine AI runs on is part of the job. We design it, source it and install it: a compact edge node in the site office, an NPU or GPU server in your rack, an air-gapped build for records that must never touch a network. Sizing starts from the workload: the documents, models, users and response times involved — never from a chip count.

Specification starts from the workload: model, concurrency, latency and data boundary before any hardware is chosen
Sourcing through established enterprise technology supply chains
Deployment into office, site, rack or isolated environments

See the infrastructure practice →

Inference architecture

Put each workload where its risk belongs.

Most businesses need more than one answer. We design the operating model that spans them — isolated runtimes for the sensitive work, approved cloud services where you allow them, with routing, policy controls and audit logging in between.

Tiering by data sensitivity and failure cost
Policy-controlled routes between local and cloud
Cost, quality and model-behaviour monitoring

Inference topology

Four places AI can run — starting with hardware you control.

Most operations use more than one tier. Placement is decided per workload, and the routes between tiers are policy-controlled and logged.

The base offering is air-gapped and on-premise AI, run on your own hardware. Sovereign cloud covers the work where local hardware isn't warranted. Approved cloud is the exception — used only with your say-so, every route logged and revocable. And for any workload you can tell a reviewer, on the spot: where it runs, what it can reach, who approved what it did.

Already running Microsoft Copilot? Then the approved-cloud tier is in place. What remains is the workloads it shouldn't touch — pricing, safety files, contract terms — and the evidence trail when AI acts. We build the local tiers those workloads need and put the controls across both.

Air-gapped

No path to the public internet. For records that physically cannot leave the building — safety files, contracts, pricing.

On-premise

Models and retrieval running on hardware you control, inside your own network, with no third-party processing.

Sovereign cloud

Australian-hosted infrastructure under Australian law, where local hardware is not warranted but jurisdiction still matters.

Approved public cloud

Microsoft Copilot, OpenAI and Anthropic models, reached through policy-controlled routes. Used only where you approve it, for work cleared to leave your environment — logged and revocable.

The air-gapped and on-premise tiers can run on hardware we design, source and deploy for you — or on airon, our appliance in development.

Start with the work

Bring us the workflow that is costing you.

The first conversation is about the work, not the technology. If an audit is the right next step, we scope one. If it is not, we say so.