JevWorks: Self hosted decision model playing a procedural game live (github.com)

JevWorks has launched an innovative self-hosted decision model that plays a procedural game live, utilizing two distinct decision-making models hosted on Hopsworks. The models, Qwen3-4B with SemIf and NVIDIA Kumo Tabular, respond to game situations in approximately 30 milliseconds, determining the best moves for a character in a dynamically generated environment. Qwen reads game scenarios and evaluates each possible move using a single forward pass, while Kumo learns from labeled game contexts without requiring a training phase. This setup showcases real-time decision-making capabilities in a complex and changing landscape. This development is significant for the AI/ML community as it highlights the potential for advanced model deployment that allows for continuous learning and adaptation in real-time scenarios. The procedural nature of the game, where obstacles and layouts are generated uniquely each time, presents a challenging testbed for AI decision-making. Despite Qwen achieving a notable single-run best distance, Kumo demonstrated superior consistency and accuracy, accomplishing a median run distance that is double that of Qwen. This project not only furthers the understanding of model performance in dynamic environments but also emphasizes the importance of both model architecture and deployment strategies in practical applications.