A defence customer requires everything on-premises with no internet. How does your architecture change?
Think first
Model choice, hardware, updates, evaluation and operations all change.
Strong answer
Use open-weight models sized to the hardware they can provide, likely quantised, and accept a capability gap versus frontier APIs, so narrow the use case and evaluate against it early. Everything the cloud gave you must now be on-site: model serving, vector store, embedding model, observability, and an evaluation harness. Plan the update path: how new model weights, security patches and index rebuilds cross the air gap and get approved. Identity integrates with their directory; logging goes to their SIEM. Capacity planning matters because you cannot burst. Budget time for accreditation. Set expectations in writing about quality and latency trade-offs.
Weak-answer indicators
- Assumes a frontier API
- No plan for updates across the air gap
- Ignores capacity planning
Likely follow-up questions
- How would you evaluate a new model version before it crosses the air gap?