AMD-Track1-Agent

This submission is a token-efficient routing agent built for AMD Hackathon Track 1, where scoring first gates every answer on accuracy and only then ranks by total tokens spent. Rather than sending every task to one large model, the agent first classifies each prompt locally with zero-cost regex heuristics into one of eight capability categories (math, logic, code debugging, code generation, summarization, sentiment, NER, factual knowledge), then routes it to the cheapest Fireworks model proven to hold accuracy for that category: MiniMax M3 for mechanical, low-reasoning tasks, and a Kimi K2 coding-specialized variant for everything requiring real reasoning, including non-code categories. That last choice wasn't assumed — it came from head-to-head testing that caught a real failure mode: the generalist Kimi model silently dumped raw chain-of-thought and never produced a valid answer on logical-reasoning prompts, a near-certain accuracy-gate failure that a coder-tuned sibling model at the same price avoided entirely. The container is a stateless batch job: read tasks, classify, call one model per task with retry-and-fallback, write results, exit — verified against the real container contract and stress-tested up to 1200 concurrent requests with zero failures.