Adaptive Routing Agent

Our AMD Developer Hackathon Track 1 entry is a general-purpose AI agent, packaged as a Docker image, that answers natural-language tasks across eight domains: factual knowledge, multi-step math, sentiment classification, summarization, named-entity recognition, code debugging, logical deduction, and code generation. Submissions must first pass an LLM-judge accuracy gate; those that pass are then ranked by the fewest total tokens used, so token efficiency, not raw accuracy, is what wins. Our design is a hybrid router. Each task is first classified almost for free with layered regular expressions, falling back to a tiny two-token model call only when the intent is genuinely ambiguous. The router then sends the task to the smallest Fireworks AI model that reliably clears the accuracy bar for that category: lightweight models handle sentiment, NER, and short summaries, while larger models are reserved for math, logic, and code. Model tiers are derived at runtime from the allowed-models list, so nothing is hardcoded. Prompts are deliberately terse, output length is capped per category, and every call runs concurrently through the judging proxy under a strict internal deadline, so valid results are always written on time.