From passive cameras to an intelligent production hall observer

The client relied on extensive camera coverage across the production hall but lacked the capacity to actively monitor feeds or efficiently analyse incidents after they occurred, leading to delayed reactions and incomplete insight.
ARP Ideas delivered an AI-based process observer that automatically analysed camera snapshots, generated event logs, and triggered alerts, enabling faster responses and clearer operational visibility.

Before the project, cameras were already installed across the production hall, but their value was limited by the need for human attention, as operators and managers could not realistically monitor more than a fraction of the available feeds or spend hours reviewing recordings after an incident.
Issues such as material leaks, sparks, manual operator intervention, or an empty tape often became visible only after downtime, quality issues, or safety concerns had already escalated.

ARP Ideas proposed a different approach by shifting from continuous video analysis to lightweight snapshot-based observation, where images were captured every few seconds at the edge and sent for analysis instead of streaming full recordings.
Each snapshot was processed by a vision-based AI model that described the situation in clear operational language, such as normal operation, service panel access, spilled material, or a stopped conveyor belt.

These descriptions were combined by an event description agent into a chronological timeline covering an entire shift or day, creating a concise and readable operational log that replaced hours of manual video review.
Instead of scrolling through eight hours of footage, a shift manager could review a short, structured list of key events and understand when downtime occurred, what actions were taken, and where problems originated.

In situations that posed immediate risks, such as health and safety hazards or unplanned line stoppages, the system automatically sent alerts to the maintenance duty officer via email or Microsoft Teams, ensuring faster intervention.
Each camera could be calibrated individually so the model understood what normal operation looked like in a specific location, improving anomaly detection accuracy over time.

From an operational perspective, the solution was positioned explicitly as a digital process observer rather than an employee surveillance tool, focusing on machines, flows, and anomalies rather than individual performance.
The architecture was designed for gradual adoption, allowing a pilot to start with a single camera in a critical area before scaling to additional lines, with Azure Storage used for snapshot handling and vision models providing the analytical layer, potentially supported by Dynamics 365 or Copilot-based summaries for supervisors.