.png&w=828&q=75)
TokenSage is a token-efficient routing agent built for Track 1 of the AMD Developer Hackathon: ACT II. It uses a lightweight, zero-token heuristic classifier to determine task complexity in microseconds. Simple/factual tasks are handled by the local Ollama model (zero cost), while complex or multi-step tasks are escalated to Fireworks AI only when necessary. The router employs a multi-stage pipeline: prompt normalization + answer caching, complexity pre-screening, local-first drafting with self-verification, and remote fallback only on failure. Only Fireworks tokens count toward the leaderboard score β every local response is free. The result is maximal local routing rate with accuracy held above threshold, proving that routing intelligence wins over raw compute.
13 Jul 2026