Infrastructure · Public sector
Local LLM infrastructure for the public sector
In public institutions the blocker is usually not the data but the inability to show where it goes. Zero-Shot builds infrastructure that runs inside the institution's network without an outbound connection: models run on institutional hardware, every access is logged and retention rules are written before anything is deployed.
What we build
- 01Model deployment inside the institution's network, with no internet dependency
- 02An inference layer for internal document search and summarisation
- 03Role-based access, quotas and complete access logging
- 04Hardware sizing and a capacity plan for concurrent users
- 05Handover to the institution's team with operating documentation and training
How it runs
01
Scope and regulation
Which data classes enter the system and how long anything is retained is written down. Deployment starts after that document exists.
02
Capacity plan
Hardware is calculated from concurrent user counts. Too much is waste; too little is a system nobody uses.
03
Closed network deployment
The system is brought up without outbound access, with an update path designed for a closed network.
04
Handover and training
The institution's team operates the system themselves. Leaving no dependency is part of the delivery.
What we measure
- Concurrent user capacity and p95 latency
- Access log completeness
- How independently the internal team operates the system
- Usage intensity — an unused system is a failed system
- Hardware utilisation and unit cost
When this is the wrong fit
Setups that make automatic decisions about citizens, or that could cause loss of rights, are outside this scope. The system is built for document access, summarisation and internal efficiency, with the decision staying human.
Common questions
Does it work without an internet connection?
Yes, it can be deployed for a closed network. Model weights are loaded onto institutional hardware and no outbound call is made at runtime. Updates follow a controlled, logged path.
What hardware is required?
It depends on concurrent users and the chosen model size. The capacity plan is produced before deployment, aiming for a configuration slightly above need but not idle — calculated from real usage estimates rather than a vendor list.
Can we see who did what inside the institution?
Yes. Role-based access and full logging are standard. Which user reached which document and ran which query is recorded, with retention set by your own policy.
Other services
What we've written on this
- Why AI agent projects failAgents that impress in a demo stall in production. Four observations on tool design, error handling, handing off to a human, and the cost ceiling.
- Is it safe to send company data to AIWhat happens when company data reaches a language model: provider terms, training use, data protection law, cross-border transfer and self-hosting.
- Making LLM costs predictableWhy AI spend surprises you at month end, and how to control it: model tiering, prompt caching, output length, and per-request ceilings.
Last updated: 2026-09-05
Let's talk about this one
Tell us what you want to build and within two days we'll tell you who takes it and how we'd measure it.