FireRoute: Token-Efficient Multi-Model AI Agent

FireRoute is a smart-routing AI agent built for the AMD Developer Hackathon's Track 1 challenge: answer natural-language tasks across 8 capability categories (factual knowledge, math reasoning, sentiment classification, summarization, named entity recognition, code debugging, logical reasoning, and code generation) while minimizing paid API token usage. Our approach: a lightweight local Qwen2.5-3B model (quantized, running entirely on-device) handles pattern-friendly tasks like sentiment analysis, NER, and summarization for zero token cost. Tasks requiring deeper reasoning — math, logic puzzles, and code — route to specialized Fireworks-hosted models: kimi-k2p7-code for coding tasks, and minimax-m3 for everything else. A confidence check automatically escalates any weak local-model answer to Fireworks as a safety net. We also disable hidden "thinking" tokens on reasoning-capable Fireworks models via the reasoning_effort parameter, cutting output token usage without sacrificing answer quality — verified firsthand after catching a code-generation task returning hundreds of tokens of internal reasoning before its actual answer. Built with Docker, llama-cpp-python for local inference, and the Fireworks OpenAI-compatible API.