Manufacturing AI at Scale: Slashing Agent Latency and Cutting 74% of LLM Routing Calls in Industrial Safety Dispatch

In the realm of industrial safety, reporting hazards such as equipment malfunctions, chemical leaks, and slip incidents is of utmost importance. These reports, submitted by shift workers in facilities like mines, smelting plants, and automated assembly lines, are routed through multi-agent triage systems built using LangGraph.

The problem lies in the fact that deploying general-purpose language models (LLMs) at conditional edges in these systems introduces significant latency, external API failure points, and high token costs for simple categorical decisions. This results in a delay ranging from 1 second to over 30 seconds, which can be critical during emergency equipment shutdowns or other time-sensitive situations.

The solution proposed is confidence-gated routing with the System 1 decision model, TypeSafe's Jev. Jev evaluates incoming safety reports in approximately 82 milliseconds. If confident, it autonomously dispatches the ticket; otherwise, it falls back to the existing LLM router without disrupting the system. By benchmarking this approach on 425 real-world industrial plant accident reports, it was found that 74% of LLM calls were eliminated, decision latency was reduced to 82 milliseconds, and net routing costs were cut by 61.3%.

The architecture also includes a non-invasive Shadow Mode, which validates agreement against the existing router on live plant traffic before making any changes in production. This ensures a zero-risk rollout of the new system.