Running an LLM inside an air-gapped facility

Industrial facilities with confidential process data — semiconductor fabs above all — do not let telemetry leave the building. Any AI that requires a cloud round-trip is disqualified before evaluation starts. That constraint shaped one of the core engineering decisions in Agent Athena™: the reasoning layer runs entirely on site.

Running a language model inside an air gap is a different discipline from calling an API. Model selection is constrained by what a single on-site GPU server can serve. Updates arrive as signed offline packages. And the model's job must be scoped to what a local model does well — structured reasoning over evidence the detection layer has already produced, not open-ended generation.

This article covers how we benchmarked local models across engineering scenarios, what the selection criteria were, where local models are genuinely sufficient, and where we deliberately keep them out of the loop.

[Full finding in preparation — Ali Taheri.]