Private AI platform build
A private language model that answers questions over your own documents, running in your own cloud or data centre.
Book an assessmentWhat you get
A private AI platform is a private LLM deployment that your organisation owns: the search over your documents and the infrastructure that runs it, inside your estate.
Design you can assure
Architecture and design documentation, high level and low level, written for your security and compliance teams to review.
A private model over your documents
A language model with retrieval over your own documents, running in your cloud tenancy, your data centre, or both.
Code, pipelines and handover
Infrastructure as code, deployment pipelines and a documented handover, so every environment can be rebuilt from your repository and audited.
The code, configuration and documentation are yours. We supply no applications, data or services of our own: we use only services you provide, or that you allow us to buy on your behalf, and every licence, account and detail is in your organisation's name, so nothing locks you in to us.
Scope, decisions and what you receive
In scope
- Architecture and design documentation, high level and low level.
- Ingestion from the document sources you agree, embeddings and a vector index.
- The model, hosted the way you choose, and the GPU compute it runs on.
- Single sign on, permission aware retrieval and private networking.
- Logging into your tools, the egress allow list, infrastructure as code and deployment pipelines.
- Acceptance testing and handover.
Decisions required
What your organisation decides, with our advice:
- Where it runs: on premises, in your cloud tenancy, or both.
- How the model is hosted: open weight on compute you control, or a managed model in your own subscription.
- Which document sources it reads, and which groups may see which collections.
- Whether prompts, answers and conversation history are kept, and for how long.
- What may cross the boundary: the updates route, any telemetry, support access.
- Who runs the platform after handover.
Delivered
- The architecture pack (see what it contains).
- The platform, running in your estate and passing the agreed acceptance checks.
- Code and pipelines in your own source control.
- Runbooks for operation, restore and incidents.
- A signed handover, with a recorded rebuild and our access removed where these are agreed in the scope.
Not included: hardware, software or services of our own. Anything the platform needs is bought in your organisation's name, by you, or by us on your behalf if you allow it.
Scope and fee are agreed in writing before any work starts.
On premises, in your cloud, or both
Where the platform runs is the first decision, and it shapes everything after it.
On premises
GPU servers in your own data centre, on your existing virtualisation (VMware or Proxmox) or on Kubernetes. The most control, and you carry the hardware.
In your cloud tenancy
GPU machines or Kubernetes in your own Azure or Oracle Cloud subscription, in a UK region, reached only over private networking. Capacity without buying hardware, in a subscription you control.
Hybrid
Documents and the vector database stay on premises while inference runs on GPU capacity in your cloud tenancy, or the other way round, joined by private links. It suits estates that already span both.
The right choice depends on how many people will use it, how sensitive the documents are, the estate you already run and what your security team will accept. An AI readiness assessment sets these choices out for your organisation before you commit. If you want the technical background first, read what it takes to run an LLM on premises.
Private document AI: RAG over your own documents
Most of the value of a private LLM comes from what it can read. Retrieval augmented generation, or RAG, finds the passages in your own documents that answer a question and passes them to the model with it, so answers draw on your material rather than on whatever the model learned in training.
In a private platform every part of that pipeline sits inside your estate:
- Ingestion. Documents are collected from the stores you choose, such as file shares, SharePoint or document management systems, and split into passages.
- Embeddings and the vector database. Each passage is turned into an embedding and stored in a vector database that runs in your environment.
- Retrieval and answer. The question is matched against the vector database, and the model writes an answer from the passages it finds.
One access rule matters most in regulated work: people only get answers drawn from documents they are already allowed to open, so the platform never becomes a way round your permissions. Where a source system exposes its permissions, they are carried through to retrieval. Where it does not, documents are split into separate collections, one for each audience. Tests with accounts that should be refused are part of acceptance. Whether answers point back to the passages they came from, so people can check them, is agreed in the scope.
Using Azure without sending data to a public AI service
Many regulated organisations already run on Microsoft, and Azure is often the quickest route to a private AI platform. There are two ways to host the model, and they are not the same, so we describe them plainly.
A model on compute you control
An open weight model runs on GPU machines in your own Azure subscription, or on your own servers. Documents, prompts and answers are processed only on infrastructure you run. No model provider receives them.
A managed model in your own subscription
An Azure OpenAI model is deployed in your own Azure subscription, as a regional pay as you go or provisioned deployment in a UK data centre, and reached only through a private endpoint with public network access switched off. Prompts and answers are processed in the UK region you chose, and your data stays in your tenancy. Microsoft runs the model under your agreement with Microsoft.
Two settings need a deliberate decision. Microsoft states that prompts and outputs are not used to train its models, but its abuse monitoring can keep prompts for review unless an exemption is approved. And Global and Data Zone deployments, which often reach the newest frontier models first, can process prompts in other regions such as the EU or the US. We use them only if your security team chooses to allow it, and we set out both settings in the design so that choice is made with the facts in front of them.
Neither option is a public AI service: a consumer chat tool used over the internet on the provider's terms, which we do not use. To understand the difference between where data is stored and who controls the system, read what sovereign AI means.
If you already have Microsoft 365 Copilot
A private platform differs where Copilot is not designed to reach:
- Model choice. You choose the model and when it changes; Copilot uses the models Microsoft selects.
- Data outside Microsoft 365. Document management systems, research stores and databases are read where they are, without first being connected into Microsoft's service.
- Collaborator collections. Partners and collaborators can be given their own scoped collections, separate from your tenancy's content.
- Logging. You decide whether prompts and answers are logged in full, as metadata or not at all, and where those logs live.
- Work not approved for Copilot. Material your organisation has kept out of Copilot can have its own boundary and its own rules.
Many organisations use both: Copilot for everyday Microsoft 365 work, and a private platform for the material that needs its own boundary.
Built for your security team to review
A private AI platform is only useful if your security and compliance colleagues can sign it off. We build for that review from the first design.
Zero Trust access
Every request is authenticated against your identity provider, and every component is reached over private networking, with no public endpoints unless one is agreed and recorded in the design.
Logging and change
What is logged, and where, is agreed in the scope: access, administration and model use can be logged where your team already looks, with changes going through your own change process.
Standards
We design to the CIS Benchmarks and Critical Security Controls, the NCSC Cloud Security Principles and the Microsoft cloud security benchmark.
Documentation for your DPIA
The architecture documents can describe data flows and trust boundaries in the terms your data protection officer needs for a data protection impact assessment, as agreed in the scope.
Read how a private platform stays in your control, section by section: the data boundary, identity, logging, the operating model and handover.
How the build runs
- 01
Agree the boundary
We agree what the platform must do, which documents it may read, who may use it and where it will run.
- 02
Design and document
The architecture is written up before anything is built, so your security team can review it early.
- 03
Build as code
Infrastructure and pipelines are defined in code and deployed into your environment through your change process.
- 04
Load and test
Your documents are loaded, and answer quality is tested against real questions from the people who will use it.
- 05
Hand over or run together
We operate the platform alongside your engineers or hand it over fully documented.
If you want senior oversight to continue after handover, our AI Solution Architect service provides ongoing architecture oversight.
For teams moving a working prototype into their own estate, the route from prototype to operated platform shows which decisions come first.
Where to go next
Start with a readiness assessment
Find out which documents and use cases are safe to start with.
Prove one use case first
Build one application on your own data and measure it before a full build.
Private AI questions answered
Costs, hardware, accuracy and UK GDPR.
Assurance for your reviewers
What stays inside your estate, and what your security team receives.
Start with an assessment.
Not sure which hosting option fits? An assessment sets out the choices for your data and your estate.
Thank you. Your request is on its way. We will reply from info@tlgsolutions.co.uk to arrange a first discussion.
That did not send. Please email info@tlgsolutions.co.uk instead, and we will reply from there.