Operating Scalable AI for Games: From LLM Intelligence to Real-Time ML Systems

Cheating in live games does more than break the rules. It damages fairness, player trust, and the health of the service itself. In this session, KRAFTON shares how we built scalable AI systems for PUBG anti-cheat by connecting two signals: what players say outside the game and what suspicious players do inside it.

We first show how noisy community feedback across languages, slang, and sarcasm can be turned into a clear operational signal using Vector Search, LLM validation, and LLM-as-a-Judge evaluation. Then we move inside the game and share how a lean engineering team evolved an ESP cheat detection system from research code into near-real-time ML pipeline on Databricks. Along the way, we discuss the practical tradeoffs behind latency, cost reduction, feature consistency, serving stability, and scale-out architecture.