Altered State Machine × FIFA · 2023
FIFA: AI League
Overview
FIFA: AI League is a mobile football game developed in collaboration with FIFA, built around a machine learning–driven agent system running entirely on-device. The core technical challenge was enabling real-time inference across multiple simultaneous agents on low-end Android hardware without server-side processing.
I integrated TensorFlow-built models into the Unity runtime, achieving 16 model inferences per second across eight agents while maintaining ~75 FPS on older Android devices. Model outputs fed into a rule-based decision layer that selected actions based on current game state, combining the predictive power of ML with the reliability of explicit logic. I also designed and built the training infrastructure — Unity ML-Agents extensions for custom agent configurations, containerised EC2 training sessions running parameterised runs to generate model variants at scale, and analytics pipelines to automate model evaluation and selection for integration into mobile builds.
Media
Key Contributions
TensorFlow model integration into Unity runtime for on-device multi-agent inference
16 model inferences per second across 8 simultaneous agents at ~75 FPS on older Android hardware
Hybrid ML + utility AI decision system combining model outputs with rule-based state logic
Unity ML-Agents extensions for custom model configurations and agent behaviours
Containerised EC2 training pipelines with parameterised runs generating model variants at scale
Analytics-driven automated model selection integrated into the CI/CD process