Agentic AI automation of UNICOS application specification processes

This project is part of the CERN OpenLab summer student programme, hosted by the BE-ICS (Beams Department, Industrial Control Systems) group. It falls under the broader Industrial Edge-Cloud and AI-based Agents initiative, a collaboration between CERN and Siemens exploring how AI can integrate with industrial control systems.

We built an AI-powered copilot agent to automate time-consuming and error-prone tasks in the UNICOS (UNified Industrial Control System) application development process. UNICOS is CERN’s framework for developing PLC-based control applications used in cryogenics, cooling, HVAC, and other continuous processes across the accelerator complex.

Project Goal

The core objective was to build an AI-powered copilot agent that assists expert users in writing comprehensive and accurate Functional Analysis documents and specification files. The agent performs two key functions:

1. Functional Analysis assistance: helps users draft project specifications by checking for completeness and consistency, proactively requesting missing details, and ensuring a well-defined project scope.
2. Specification transformation: automatically converts user-written textual Functional Analysis into the structured spreadsheet format required by UNICOS for automated PLC application generation.

This bridges the gap between high-level requirements and the structured specification needed for development.

Approach

We built an MCP (Model Context Protocol) server that integrates directly with the UNICOS development workflow. The server exposes tools to an AI agent capable of:

• Analyzing user-provided text to identify missing information and inconsistencies.
• Extracting key requirements and translating them into structured data.
• Generating suggestions and prompts to guide users toward complete specifications.

The MCP server architecture keeps the AI logic separate from the UNICOS tooling, making the system modular and framework-agnostic.

Roles

Daniya is leading the implementation:

• Designing and implementing the MCP server in Python: architecture, tool definitions, and API integration.
• Working with CERN experts to understand the Functional Analysis contents and the UNICOS specification file format.
• Developing the conversion logic from natural language requirements to structured spreadsheet cells.
• Testing and evaluating accuracy, reliability, and usability of the assistant.

I’m supervising the project, providing guidance on the MCP architecture, AI agent design, and integration with the existing UNICOS ecosystem.

This project gives BE-ICS a hands-on experiment with AI-driven automation in industrial control systems, with the potential to significantly reduce specification errors and accelerate project delivery.

Results and Conclusions

The implemented prototype demonstrates that LLM-based agents can be integrated into the existing UNICOS engineering workflow without replacing established tools. Key results include:

• FA compliance checking: The system extracts verifiable rules from reference templates, performs bidirectional section alignment using hybrid retrieval, and produces compliance reports comparing project documentation against requirements.
• Specification generation and validation: The agent orchestrates existing CoVeET and SpecsBuilder tools through MCP, validates generated workbooks against FA evidence, SpecDoc constraints, and UNICOS documentation, and performs constraint-aware editing that rejects invalid modifications.
• Hybrid retrieval: Combining semantic and keyword search handles cases where project and reference documents use different section structures, enabling robust document matching.

The project establishes the feasibility of augmenting UNICOS engineering tools with LLM-based document interpretation and reasoning. However, systematic evaluation against engineer-verified cases is required before considering operational deployment.