Live-Service Anomaly Detection
A production AI investigation system for PB-scale game telemetry, monitoring 50+ service KPIs and shortening analyst response latency by more than 40%.
- telemetry analyzed
- PB-scale
- live-service KPIs
- 50+
- lower response latency
- 40%+
Project note
At Activision Blizzard, the practical challenge was not simply detecting an unusual metric. Analysts needed to understand which live service changed, gather relevant context, and move from an alert to a useful investigation while the signal was still actionable.
The resulting workflow processed PB-scale telemetry in Databricks and Azure and monitored more than 50 service KPIs. Anomaly signals were connected to a research layer combining retrieval, deep-research patterns, LangGraph orchestration, and a Slack-based MCP agent. That reduced analyst response latency by more than 40%.
Airflow and GitHub Actions moved the system from a proof of concept to daily automated reporting. The work was the only intern project selected for the 2025 Microsoft Xbox Game Studios Data & Applied Science Summit. This public case study intentionally omits proprietary data, model settings, and internal service details.