At Orange Morocco, network issues that slip past monitoring turn into customer complaints. I built the real-time pipeline that closes that gap – processing more than 4 million records a day of telecom activity and turning it into early warnings.
Modular PySpark jobs orchestrated by Airflow clean and aggregate the stream, while layered anomaly detection (Isolation Forest, Autoencoders, and time-series methods) catches abnormal behavior as it emerges. A unified data lake replaced fragmented sources, and curated SQL layers feed live Power BI dashboards through REST APIs.
The impact was concrete: KPI latency dropped from hours to minutes, and the system caught 25% more incidents before they reached customers, with MLOps in place to keep the models fresh.