KRAFTON at Scale: Architecting Real-Time Game AI with MLflow & Serving

This session explores KRAFTON's journey of deploying real-time AI for PUBG: BATTLEGROUNDS, a title with 200M+ users. We share how a lean engineering team overcame global-scale challenges using Databricks. 1. Real-time anti-cheat MLOps: We detail the transition from an hourly batch system to a sub-minute real-time pipeline using Structured Streaming and an online feature store. By optimizing Model Serving endpoints, we drastically reduced latency and slashed costs to 1/10th, enabling just three engineers to support a massive global player base. 2. Esports win prediction: We discuss the serving strategy for Cox PH models to handle 64-player dynamics. Solutions include sequence buffering to ensure data ordering and incremental feature engineering for O(1) efficiency, delivering stable predictions with sub-200ms latency.We conclude with Databricks MLOps patterns bridging the gap between research and production, alongside our roadmap for next-gen VLM-based anti-cheat.