Track 08 · Deployment & opsLesson 05 of 11
Self-hosted, VPC and air-gapped
Run models where regulated customers require it, and know what it costs.
- Specialisation · Government
- 35 min read
- 75 min hands-on
- Pro
What you'll be able to do
01- Place a customer requirement on the deployment spectrum and explain what crosses the boundary at each step.
- Configure private connectivity to a managed model API on AWS, Azure or Google Cloud, and explain why a private path is not residency or self-hosting.
- Design a self-hosted inference tier on open weights with vLLM or SGLang: local weights, telemetry off, eval re-run before promotion.
- Implement an air-gap import path (signed manifest, verification, mirrored registries) and an egress check that fails on any external endpoint.
- Design support for a system you cannot see: diagnostic bundles, runbooks, offline licensing, patch cadence and secrets.
Before you start
- Secrets, dependencies and network isolation: signed images, SBOMs, default-deny egress, vaulted secrets. This file asks what changes when nothing can be fetched.
- Data residency and sovereignty: why a Region, country or operator is required.
- Deploying into the customer's cloud, Containers and Kubernetes for agents, Infrastructure as code for agents and CI/CD and the agent lifecycle.
- Choosing a model you can defend and Regression tests and CI gates.
- Background: DNS resolution, CIDR ranges, how a container runtime resolves an image reference.
The short version
Deployments sit on a spectrum from vendor SaaS to fully air-gapped. Each step moves the agent, the data and eventually the model inside the customer's boundary, and moves operational work from you to them. Private connectivity (PrivateLink, Private Endpoints, Private Service Connect) keeps traffic to a managed model off the internet, but a provider still processes the prompt. Self-hosting open weights keeps the model inside, at the price of owning capacity, quality and patching. Air-gapped means nothing is fetched at runtime: every artefact arrives through a signed, verified import path, nothing phones home, and you support the system through diagnostic bundles and runbooks.
- Choose the topology from the customer's written hard requirement.
- Private path, residency and self-hosting are three different promises.
- In an air gap, the import pipeline is the supply chain.
- Anything that phones home is a defect.
- You cannot see the system, so make it diagnosable by someone else.
Pro lesson
The rest of this lesson is in Pro.
The mental model, worked example, field note, failure table, exercise and self-check, with the voice tutor.
- Every stage and specialisation, in full
- Every lab and every role project
- Halden weeks 2–10, on your record
- Every drill, live in the playground
- All solutions, plus the server judge
- Lab Assistant and every Teach with AI