William Heuser & Sofus Erdal

From One Noisy Model to Specialized Agent-Orchestrated Models

Predictive maintenance systems often fail in production due to noisy data, weak labels, and the high operational cost of false positives. In this talk, we present our shift from a single monolithic model for EV charger maintenance to a decomposed architecture of specialized models, each focused on a specific failure mode with a clear action path. We then introduce an orchestration agent that converts model outputs into operational workflows—opening and updating tickets, executing low-risk remediation, and learning from incident outcomes. The result is a closed-loop AI system that reduces noise, accelerates resolution time, and builds operational trust.

Bio: Founded by a team from the University of Copenhagen with backgrounds in Computer Science, Economics, Mathematical Modelling, and Business Analytics, we combine strong technical and analytical foundations with practical industry experience. Today, we monitor 40,000 EV chargers across Europe, scaling rapidly through an AI-first architecture and extensive use of advanced AI models across our platform.