WarriorsPersonalAssistant

Our hybrid routing system revolutionizes function calling by intelligently balancing on-device FunctionGemma execution with cloud Gemini fallbacks, achieving optimal performance through adaptive traffic shifting, query normalization, and deterministic argument extraction. The architecture implements sophisticated agentic workflows that prioritize local-first reasoning for speed and privacy, utilizing FunctionGemma and Cactus Compute to redefine the edge-cloud frontier. Our model persistence optimization eliminates redundant initialization overhead, while the adaptive traffic shifter learns optimal routing patterns per query category through exponential moving averages. Query rewriting normalizes indirect phrasing before processing, and hybrid argument extraction separates tool selection (FunctionGemma) from argument parsing (regex), ensuring both accuracy and efficiency. This multi-layered approach dynamically escalates complex queries to cloud processing while maintaining high on-device ratios for simple requests, fundamentally transforming how AI systems balance local computation with cloud resources.

This project was developed entirely during the hackathon period. The core hybrid routing architecture, adaptive traffic shifter, model persistence optimization, and regex argument extraction systems were all designed and implemented from scratch. While we utilized existing APIs (FunctionGemma via Cactus Compute and Gemini 2.5 Flash), the intelligent routing logic, multi-phase execution strategy, and agentic workflow coordination represent novel contributions that fundamentally advance local-first function calling capabilities.