SEVK: Saving Expenditure Via Knowledge

This project is a sophisticated submission for the AMD Developer Hackathon: ACT II (Track 1), built around a highly optimized hybrid routing agent. The primary objective of the system is to process diverse datasets while aggressively minimizing Fireworks API token consumption. To accomplish this, the agent employs an intelligent routing mechanism that evaluates the nature of incoming tasks before execution. For structured and deterministic challenges, such as mathematical calculations or logical puzzles, the system routes the query entirely away from the API to zero-token local solvers, saving significant computational and token resources. For tasks that require natural language understanding and must be sent to the API, the agent utilizes advanced optimization strategies to ensure minimal token waste. This includes a strict cross-tier fallback mechanism and aggressive token-starvation tactics, such as disabling reasoning tokens and stripping out unnecessary tags. Furthermore, the system includes sophisticated bundling logic that aggregates compatible tasks together, significantly reducing the overhead of multiple distinct API queries. The entire evaluation pipeline is fully containerized using Docker, ensuring a seamless, plug-and-play experience for testing and deployment. In addition to the core evaluation agent, the project features an interactive Streamlit web application. This local UI serves as a demonstration tool, providing a real-time, side-by-side comparison of the fine-tuned token-efficient router against a standard prompt-based baseline, clearly showcasing the token savings and efficiency gains achieved by the system's architecture.