Private AI, inside your own boundary.

01 / Private AI: running language models inside your boundary

What private AI means.

Four components make up a working document-and-question system: the model runtime, the retrieval index, the document store and the audit log. Private AI is the arrangement where all four live inside one boundary the organisation controls, on its own premises or in a dedicated environment built for it, and nothing leaves that boundary to be processed elsewhere.

The word "private" is doing real work. It is not a model hosted by a vendor with a privacy policy. It is infrastructure the organisation can inspect, with a list of permitted crossings written down before the first document is indexed.

02 / Private AI: running language models inside your boundary

When it is required rather than preferred.

Residency rules that forbid processing outside a jurisdiction. Confidentiality obligations to clients or patients that a third-party terms page cannot discharge. Regulators who will ask to see the log. Material that must never be used to train anyone else's model. In each case the question is not whether the vendor is trustworthy but whether the organisation can demonstrate control. A private deployment lets it.

03 / Private AI: running language models inside your boundary

What runs inside the boundary.

The runtime serves the model. The index turns documents into something the model can search. The store holds the documents and their versions. The log records every question, the sources retrieved, the answer, the model version and who approved it. Around them sit the identity service users sign in with, the audited path administrators use, and the monitoring that says whether the system is healthy.

Residency is a design decision: where documents may be stored, where processing may happen and which jurisdictions are excluded are recorded as requirements and then met by placement and routing. Changing that list goes through the same change process as changing the model.

• Model runtime, retrieval index, document store and audit log inside one boundary
• Users through the organisation's identity service; administrators through an audited path
• Residency and permitted outside services written down and enforced by design
• Client documents, questions and answers never used to train a model unless asked for in writing

04 / Private AI: running language models inside your boundary

Hardware and models, honestly.

The hardware depends on the models and the workload: how many people ask questions, how long the documents are, how fast an answer must come back. Memory on the accelerator is usually the constraint that decides which models are practical. We state the sizing for a particular environment after seeing the workload, not from a table, and we would be suspicious of anyone who does otherwise.

Open-weight models run entirely inside the boundary and are the natural fit for private deployment. Where a controlled commercial model is justified, it is named in the design together with its version and the data that may reach it, and that is a different arrangement with its own contract.

05 / Private AI: running language models inside your boundary

Deployed is not finished.

A private system has an operator, monitoring, a version history for models and prompts, an evaluation set that is run again when anything changes, a change process, and a plan for incidents and cost. Answers cite their sources and a named reviewer approves or sends back, so the record shows what the system said and what a person decided. The organisation can operate it, be trained to, or hand it over with the records.

06 / Private AI: running language models inside your boundary

Who does this work.

Custom AI work described on this site is contracted with Genyra L.L.C-FZ in Dubai, a separate company that operates genyra.ai. The design, the infrastructure and the operating duties are discussed in an enquiry; available hardware and hosting are confirmed for the particular environment.