Local Inference for sensitive office work

Local inference runs the model on equipment in your office. We also offer self-hosted inference on a managed server, with hosting, location and access controls agreed explicitly. Data residency, operational control, network isolation and compliance depend on the design of the complete service and the controls agreed for it.

The choice depends on your documents, concurrent users, required response quality, end-to-end latency and operating constraints. An office appliance keeps inference on office equipment and requires suitable space, power, networking and maintenance. A managed server supports remote operation under an explicitly agreed location, access model and backup plan.

It can be designed to check current identity and authorisation before retrieval, blocking the request when those checks fail. Permission revocation and the cleanup of indexes or other derived copies are separate controls. Files that a user has already exported or downloaded cannot generally be recalled by the inference system. Exports can be restricted and logged, with separate handling or deletion controls for copies outside the assistant.

We test representative documents and candidate models against agreed acceptance cases. The sizing decision uses observed response quality, expected concurrent use and end-to-end latency, including OCR and retrieval where they are part of the workflow. This avoids choosing capacity from a model name or headline benchmark alone.

We manage the scope agreed for the engagement. It can include model and runtime updates, permission tests, monitoring, backup and restore work, and controlled support access. Responsibilities, access routes and maintenance procedures are documented so your team knows who handles each part of the service.