AI behind the boundary: deploying K-Lake where data cannot leave

Some organisations cannot send their documents to a cloud service under any contract. Defence suppliers, public bodies, hospitals and many financial institutions hold material that must stay inside a defined boundary, and in some cases on a network with no route to the internet. This paper explains how K-Lake brings AI retrieval to that material without asking for an exception to the rule.

Three ways to deploy

K-Lake is the same software in every configuration. What changes is where the language model runs.

In the first configuration, the organisation connects a cloud assistant it already uses. The index and permission checks stay inside the boundary, and the assistant receives only the passages retrieved for each question, under the organisation's own agreement with that vendor.

In the second, the organisation runs a model on its own infrastructure and connects it to K-Lake over MCP. No content leaves the network, although the network may still have internet access for other purposes.

In the third, the whole system runs with no external connection at all. K-Lake has been demonstrated answering questions in this mode with the internet physically disconnected.

Extraction stays inside

Text extraction, OCR and transcription run on engines the organisation hosts. That includes scanned documents, complex tables, audio and video. An external extraction service is used only if an administrator configures one, and each transfer to it is recorded with the file, the provider and the volume sent. In an isolated deployment there is nothing to configure and nothing to record.

Licensing without a call home

Software that has to contact its vendor to confirm a licence cannot run on a disconnected network. K-Lake licences are signed files covering indexed capacity and an expiry date. They are validated locally, they cannot be altered without detection, and they fail closed if they are invalid. There is no licence server and no usage telemetry. A licence can be renewed or resized without reinstalling.

Identity on a closed network

Where an isolated site has its own identity provider, K-Lake federates to it in the usual way. Where it has none, personal access tokens let named users and local agents work with the same per-file permissions and the same audit trail.

Updates and monitoring

Offline update packages carry new releases onto sites with no outbound connectivity. Every service exposes metrics and structured logs for the organisation's own monitoring and SIEM tools, so operators keep full visibility without anything being sent to Zegress.

Evidence your reviewers can gather themselves

For a regulated buyer the useful property of this design is that its claims can be tested rather than accepted. Disconnect the network and confirm the system still answers. Watch the network boundary during extraction and confirm nothing leaves. Inspect the licence and confirm it validates offline. Review the audit trail for a sample of questions.

K-Lake is designed for organisations working under DORA, FCA rules, NHS data governance and public sector data policy. Design intent is not certification, and our current assurance status is published on our trust centre and reviewed monthly.

Getting started inside the boundary

A proof of concept can run on a single virtual machine inside the target environment, using the organisation's own documents and, where required, its own model. The software and licence arrive as files. Your team installs them, so the evaluation itself follows the same rules as production.