Factory-floor AI does not need to start with cloud-first fantasy.
In real industrial environments, especially air-gapped ones, the useful pattern is more constrained:
local data
local models
governed outputs
human-in-the-loop decisions
The architecture is not “LLM replaces everything.”
It is hybrid.
Normal ML handles the things it is good at:
• defect detection
• anomaly detection
• failure prediction
• soft sensors
• forecasting
• optimization signals
The LLM layer handles the things ML alone is bad at:
• contextual reasoning
• workflow interpretation
• operator-facing explanations
• tool use
• report generation
• guided recommendations
MCP being the orchestration layer between the LLM and the local industrial context: historians, CMMS/LIMS (Computerized Maintenance / Laboratory Information..), engineering docs, procedures, databases, and approved tools.
In an air-gapped factory, this is needed because the system can't depend on internet access, cloud inference, or uncontrolled external services.
The output should not be full autonomous control.
It should be governed recommendations, explainable actions, operator approval, MES/ERP integration, and traceable reporting.
That is the practical shape of factory-floor data science:
ML for signal intelligence.
LLMs for contextual reasoning.
MCP for controlled tool and knowledge access.
Humans for authorization.
Local infrastructure for sovereignty.
Less magic.
More architecture.