AI for Control Systems: Lessons Learned from Building a Governance-First Platform

Already registered? Click here to log in. 

Summary

In the high-stakes world of industrial automation, the journey from AI theory to functional control system integration is paved with rigorous data governance. This session dives into the architectural evolution of an AI platform where the primary constraint wasn't model "intelligence," but rather the physical and legal location of data. We explore why the initial pivot to Azure Government Cloud was the catalyst for our entire design, dictating how we navigated the tension between cutting-edge LLM capabilities and the strict security requirements of on-premises and US-based hosting.  

Moving beyond the initial infrastructure, we share the lessons learned results of our experimentation with agentic frameworks and retrieval strategies. Attendees will learn why we moved away from the industry hype of multi-agent orchestration and embedding-heavy RAG in favor of deterministic state machines and hybrid search. By leveraging the Model Context Protocol (MCP) and LangGraph, we developed a portable, model-agnostic architecture that prioritizes inspectability and reliability which are essential traits for any system operating within a control environment.  

Key Takeaways:

•    Learn how data residency and sovereignty requirements (on-premises vs. US-hosted) act as the primary filter for LLM selection and hosting strategies.

•    Understand why deterministic state-machine workflows outperform autonomous agent orchestration in industrial control environments, and how LangGraph implements this pattern.

•    Why a "keyword-leaning" approach to Retrieval-Augmented Generation provides the precision necessary for technical control system documentation and metadata.

•    Discover the value of wrapping endpoints in the Model Context Protocol (MCP) to ensure your AI platform remains agnostic to specific models and easy for human operators to audit.

•    Real-world insights into why local models and complex multi-agent orchestration often fail to survive contact with actual automation projects.