The Processing Layer
The processing layer is a custom distributed inference engine optimized for sustained throughput in isolated environments — not for short bursts under ideal conditions, but for continuous operation over years. It supports both CPU and GPU execution paths, with automatic load balancing across whatever compute resources a given deployment has available.
Why No Cloud Offload
Every inference runs on hardware physically present at the site. That isn't a performance workaround — it's the architectural prerequisite for air-gap capability: a system that requires external compute to function cannot, by definition, run air-gapped. The processing layer is built from the ground up for local execution, not retrofitted for it.
From Theory to Measured Latency
The current production deployment shows what that means in practice: Henri's anomaly classification runs in under 40 milliseconds per cycle — entirely on the hardware present at the site, with no dedicated GPU infrastructure. That figure comes in production from continuous operation, not a lab benchmark.