The gap between "the model works on a laptop" and "the model runs on every PLC in the fleet" is measured in years for most industrials. We close it in quarters.
What Changes at the Edge
- Latency budgets are sub-100 ms, not sub-second
- Connectivity is intermittent — your control plane must tolerate weeks offline
- Determinism is mandatory — a probabilistic ML decision cannot replace a safety interlock
The CYBERIA Edge Pattern
- K3s clusters at the line level, fleet-managed from cloud
- OPC UA / Sparkplug for telemetry, never custom protocols
- Inference sidecars that consume telemetry and produce advisory signals — control loops stay in the PLC
- Signed model bundles with full provenance, rolled out via GitOps
Governance That Plant Managers Trust
Every inference is logged with model hash, input hash, and timestamp. When the auditor asks "why did the line slow down at 03:14?" — we have the answer in seconds.