RouteZero: Hybrid Token-Efficient AI Agent

RouteZero is engineered to dominate the AMD Developer Hackathon Track 1 leaderboard by perfectly balancing the tightrope between the Accuracy Gate and Token Efficiency. Unlike standard "probabilistic" routers that waste tokens using an LLM-as-a-judge, RouteZero utilizes a zero-cost local-first architecture. It attempts every task on a highly optimized, baked-in 4B parameter model running natively on the judges' AMD hardware. The local model's answers are then verified using lightning-fast, deterministic Python scripts. To guarantee passage through the Accuracy Gate, math and logic tasks use our custom "Agreement Guard"—sampling the local model twice and instantly escalating to the Fireworks API if the model disagrees with itself. When escalation is necessary, RouteZero routes smartly: complex math and coding tasks go to the strongest allowed models (like Kimi or Minimax), while language tasks are strictly routed to Gemma models to secure eligibility for the Best Use of Gemma sub-prize without risking overall accuracy. Finally, RouteZero constantly monitors its own Tokens-Per-Second generation speed. If it detects the judging server is running too slowly to meet the strict 10-minute deadline, it activates an emergency failsafe, seamlessly overriding to pure Fireworks API usage to prevent a timeout. The result is a bulletproof container that extracts maximum accuracy for the absolute minimum token cost.