Industrial IoT platform — 12,000 real-time sensors

12,000 sensors across 47 industrial sites, generating 8 million events/day. The existing infrastructure (on-premise, MQTT + InfluxDB) saturated at 4 million/day.

Maintenance teams wanted predictive alerts (detecting failures before breakdown) — not just supervision.

NIS2 compliance mandatory: essential operator, incident management, traceability, continuity plans.

Network constraint: several sites on intermittent cellular connections — the edge had to keep working offline.

3-tier architecture: edge (industrial Linux gateway + local MQTT broker + SQLite buffer), regional (aggregation, first rules), central (ClickHouse time-series, dashboards, ML).

Edge computing: local anomaly detection (lightweight ONNX model), immediate alerts even while disconnected, deferred sync on network return.

Predictive ML: XGBoost models trained per equipment type, data-drift detection, MLOps via MLflow, monthly retraining.

Central platform: continuous event ingestion, long-term history, real-time dashboards, and operational alerts from the first anomaly.

NIS2 compliance: asset mapping, incident management plan, geographically distributed backups, annual crisis exercises.

Fully automated, reproducible deployment — remote gateway updates included.

Ingestion capacity: 4M → 12M events/day, end-to-end p95 latency 1.8 s.

Predictive maintenance: 73% of failures detected before breakdown (vs 0 before). Estimated €2.4M/year saved in avoided downtime.

NIS2 compliance: all technical controls in place, first audit passed with a single minor gap.

Isolated sites: 100% operational even through long network outages (local buffering, replay on return).