In manufacturing, having the right tool in the right hand is a matter of both safety and quality. But how do you monitor this without risking total surveillance of your employees? Manual logs fail in the daily grind of shift work, and cloud-based video analysis often crashes against the wall of GDPR compliance.

In this session, I will show you a solution that starts right at the workbench: we use a Vision Language Model (VLM) running locally on an edge device. Instead of scanning faces, the system recognizes a simple color logic (e.g., colored gloves acting as role indicators). Images never leave the local device—only anonymous events are sent to the cloud.

What you will learn in this talk:

The VLM Workflow: I’ll walk you through the step-by-step process of how we optimized the model for tools and colors (Zero-Shot vs. Fine-Tuning). You’ll see exactly how it’s done, without us having to wait for live training progress bars.

Edge Inference & Architecture: Why the AI must run locally and how the anonymized communication with Microsoft Teams and Power BI is technically implemented.

Privacy by Design: How to convince the works council with an architecture that stores no personally identifiable information (PII).

The Live Demo: I’m bringing hardware! We will test it live on stage: Will the system detect when I reach for a drill with the "wrong" glove? From the local camera event to the alarm in Microsoft Teams—100% reproducible and practical.

This isn't a theory-heavy session; it’s a real-world blueprint for legally compliant AI in the industry.

as described in the description, will run on hardware (laptop, or dedicated thing)