Marius Ayrault

Real-Time Industrial Performance AI Assistant

An AI agent connected to the telemetry of a simulated industrial site: it answers natural-language questions about production indicators, diagnoses quality drifts and guides continuous-improvement decisions. Two worlds meeting, Lean Six Sigma operational excellence and generative AI.

  • AI Agents
  • Generative AI
  • LLM
  • Real-Time
  • Time Series
  • Industrial Performance
  • Lean 6 Sigma
  • RAG
  • Automation
  • CI/CD
  • Docker

By Marius Ayrault · 11/06/2025 · 2 min read

The problem

On a shop floor, performance indicators exist but stay locked in dashboards that few people open. The simple questions a production manager asks (how healthy is the line, which tool is degrading quality, where to focus effort) still require manual analysis.

The solution

An AI agent plugged into the real-time telemetry of a simulated production line, combining three building blocks:

  • Production simulation: an asynchronous synthetic-data generator (positions, orientations, cycle rates, breakdowns drawn from a machine-error catalog) reproduces a two-tool assembly line, one of which deliberately drifts in quality to give the agent real problems to diagnose.
  • Statistical process monitoring: a full OEE breakdown (required, operating, net and useful time), quality, performance and availability rates, MTBF and MTTR, plus Cp and Cpk capabilities computed continuously per tool on position and orientation, with distribution views and control charts.
  • Tool-augmented conversational agent: an agent built on Mistral AI (mistral-medium-2505) with a think, act, observe, respond reasoning loop, each phase streamed and inspectable in the Gradio interface. It relies on five tools: real-time production status (over 30 metrics), a downtime log with causes, a RAG knowledge base (ChromaDB, mistral-embed embeddings) over SPC literature, webpage reading through a custom web-to-markdown pipeline, and a calculator. Enough to answer both “which tool is most critical right now?” and “explain the Cp and Cpk indicators”.

Architecture: simulated assembly line, SPC statistical monitoring, tool-augmented Mistral AI agent and shop-floor actors querying in natural language

What the project demonstrates

A concrete application of my dual background: the operational-excellence methods (Lean Six Sigma, TPM, SPC) practiced at Safran, served through an AI interface that makes them accessible to everyone on the floor. With a deeper benefit: the know-how needed to interpret the indicators lives inside the agent, so it can be shared and is no longer lost to turnover. And the underlying pattern, an agent grounded in computed metrics rather than left to guess them, transfers directly to any agent system put in front of production data.

Demo

Recorded from the live application, built during the Hugging Face Agents-MCP hackathon. The source code is private.