DAB AI: Direct Action Bureau
Sovereign & Private AI you actually own.
DAB AI is a single bureau of specialists for sovereignYou own and control the hardware, data, and models, rather than renting a vendor's., private AI agents (local servers and local weightsThe model's learned parameters, stored on your own hardware. where practical), covering Direct Action, Advice, and Research.
Three desks. One bureau.
Direct Action, Advice, and Research — under one bureau standard.
Desk 01
Direct Action
We install and harden real systems — a running setup you own, not a slide deck.
- On-premRunning on servers in your own building or data centre. or private-cloudDedicated cloud infrastructure reserved for your organisation only. inferenceRunning a trained AI model to get answers — not training it.
- Agent runtimesThe software that runs an AI agent and governs its steps, tools, and limits. and tool accessWhich external actions an agent may take (read a file, call an API, etc.).
- SegmentationSplitting a network into isolated zones so a breach can't spread., monitoring, access control
Desk 02
Advice
Vendor-neutral guidance so you buy and build the right thing.
- Architecture and build-vs-buyDeciding whether to build a capability yourself or buy an existing one.
- Sovereignty and data-flow reviews
- Policy and readiness assessments
Desk 03
Research
Applied evaluation that feeds straight back into install defaults.
- Local model evaluationTesting which local models perform best for your tasks and hardware.
- Hardware sizingWorking out the compute (GPU/CPU/RAM) needed to run a model well. and agent reliability
- Public notes and private client research
Private AI you actually control.
Local servers and local weights where practical — hybridA deliberate mix — some workloads local, some in the cloud. or cloud when that is the sober choice.
Sovereignty
Infrastructure and model weights you control where it matters.
Privacy
Explicit data paths; minimise unnecessary third-party exposure.
Direct Action
We implement, not only recommend.
Candour
Local when practical; hybrid or cloud when that is the sober choice.
We will say when a fully air-gappedNo connection to the public internet; data cannot leave the system. build is right — and when it is not.
How an engagement works
From constraints to ownership — under one bureau standard.
-
Discover
- Constraints
- Threat modelA structured description of what you're protecting, and from whom.
- Success criteria
-
Design
- Stack & topologyHow your servers, services, and networks are laid out and connected.
- Model choices
- Access model
-
Install
- Servers & runtimes
- Agents
- BaselinesA known-good starting configuration you can compare changes against.
-
Harden
- Segmentation
- Monitoring & backup
- Controls
-
Hand over
- RunbooksStep-by-step instructions your team follows to run or recover a system.
- Ownership transfer
- Optional retain
Definition of done
What you have at the end of an engagement:
- Agreed services running in your environment
- Access control and operational basics in place
- Runbooks your team can follow
- Ownership transferred to you
Typical engagement patterns
How the work usually takes shape — pick the closest fit, we'll scope the rest.
Pattern A
Air-gapped inference
Private model serving with no outbound dependency for core inference paths.
Discuss this
Pattern B
On-prem agent toolchain
Agent runtimes with controlled tool access, logging, and human-in-the-loopA person reviews or approves the AI's actions at key points. gates.
Discuss this
Pattern C
Sovereignty review
Architecture and vendor review before buy-in — what must stay local, what can be hybrid.
Discuss this
What we do not do
The work we turn down — and what we do instead.
- Public-API wrappers sold as transformation — we install systems you own.
- Unaudited scraping or covert exfiltration tools — explicit, auditable data pathsThe exact route your data takes, and where it's stored or copied..
- Chatbots sold as strategic AI programmes — systems that work, owned by you.
One standard of delivery.
Meet the people who deliver under DAB AI bureau standards.
Bureau profiles
