Predictive maintenance, redesigned

The problem

The client’s predictive-maintenance engine runs about 500,000 rules, written over the years, that read sensor data and decide whether an alarm should go off. The code had become impossible to maintain, and a full run took 10 minutes on 7 heavy servers.

What I did

• Delivered the small improvements the client asked for, and in parallel designed a complete re-engineering of the engine.
• Showed the new design to the client, who then paused their other initiatives in the area to back the new MVP.
• Delivered the MVP: the same rules now run in seconds, sometimes milliseconds, on a single server, and can be tested and replayed on historical data.
• Now leading the migration roadmap and the decommissioning of the legacy system: data migration, integrations, feature completeness.
• The client’s experts now write new rules with an AI agent connected to the platform and to their knowledge sources.

Next

An AI roadmap built on top: early anomaly detection on the sensors, and generative AI to de-duplicate incidents and help with root-cause analysis.