Smart Token-Efficient AI Router

Our project tackles the core challenge of Track 1: building a general-purpose AI agent that │ balances high accuracy with extreme token efficiency. In modern enterprise environments, routing │ every query to the most capable, premium model results in massive, unnecessary API costs. Our │ solution acts as a smart gateway intercepting tasks across various categories—including factual │ Q&A, mathematical reasoning, logic puzzles, and code generation. │ │ Instead of relying on an expensive LLM-based router, we engineered a deterministic, zero-token │ local heuristic router that runs completely locally within the container constraints (4GB RAM, 2 │ vCPU). By analyzing the semantic complexity and keyword triggers in the user's prompt (such as │ coding keywords or multi-step logic requirements), the router instantly categorizes the task. │ Simple tasks are routed to a fast, low-cost model, while complex tasks that demand deep reasoning │ are elevated to a premium 70B class model via the Fireworks AI API. │ │ This hybrid approach ensures that our agent comfortably passes the strict 80% accuracy gate while │ securing a highly competitive position on the token-efficiency leaderboard. The architecture is │ containerized via Docker for seamless deployment, and relies on strict system prompts to prevent │ conversational filler, ensuring we save every possible token for maximum ROI.