The sovereignty practice
AI that never
leaves the building.
For data you can’t send offshore — client files, records, contracts, IP — we build the full capability on hardware you control: open-weight models, self-hosted, Australian-resident, and never used for training.
Why it matters
Some data should never touch a public chatbot.
The moment sensitive material goes into a free AI tool, you no longer control where it lives, who processes it, or what it trains. For businesses bound by privacy law, professional obligations or plain commercial sense, that’s not a risk — it’s a breach.
The answer isn’t avoiding AI. It’s owning it. Open-weight models are now good enough for serious work — and they’ll run in a room you can point to.
- Legal & conveyancing
- Medical & allied health
- Accounting & finance
- Engineering & IP-heavy firms
- Government-adjacent suppliers
- Anyone bound by privacy law
How it’s built
The spec sheet, plainly.
You don’t need to know this stack — that’s our job. But you should know it exists, because “private AI” often means someone else’s cloud with a padlock icon.
- Models
- Open-weight (Llama, Qwen-class), selected per task
- Inference
- Self-hosted — your hardware or a dedicated AU datacentre
- Your documents
- Private retrieval (RAG) — indexed locally, cited answers
- Sensitive paths
- No third-party APIs. None.
- Training on your data
- Never
- Ownership
- The hardware, models and system are yours
- Verification
- Auditable: we show you exactly what talks to what
The honest boundary
Where a bigger model is worth it, we say so.
Frontier cloud models are still stronger for some tasks. Good sovereignty engineering isn’t pretending otherwise — it’s designing the boundary so only non-sensitive work crosses it, and your team knows exactly which side of the line they’re on.
Most of our engagements end up hybrid: private infrastructure for the data that matters, mainstream tools for everything else, and a written policy that makes the split obvious to everyone on staff.
Point at the room. Own the machine.
The audit maps your sensitive-data paths and prices the private build — fixed, in writing.