Brocacho Hybrid Token Router

Brocacho AI Agent is a Dockerized general-purpose AI agent developed for Track 1 of the AMD Developer Hackathon: ACT II. The project is designed to balance two competing objectives: achieving high answer accuracy and minimizing the number of tokens routed through Fireworks AI. When the container starts, it reads natural-language tasks from /input/tasks.json and classifies each prompt into one of eight supported categories: factual knowledge, mathematical reasoning, sentiment classification, text summarisation, named entity recognition, code debugging, logical reasoning, and code generation. Tasks that can be solved reliably using deterministic algorithms are handled locally using Python. These local solvers perform arithmetic, percentage calculations, selected word problems, sentiment analysis, structured extraction, and other high-confidence operations without consuming Fireworks tokens. Tasks requiring broader language understanding or reasoning are grouped into compact batches. General, factual, mathematical, summarisation, NER, sentiment, and logical-reasoning tasks are processed using MiniMax M3. Code-debugging and code-generation tasks are processed using Kimi K2P7 Code. Batching allows multiple tasks to share one system instruction, reducing repeated prompt overhead. If a batch response is missing or malformed, the system uses an individual fallback request only for the affected task. The agent reads FIREWORKS_API_KEY, FIREWORKS_BASE_URL, and ALLOWED_MODELS dynamically from the evaluation environment. It never calls a model outside the permitted list and routes all remote inference through the provided Fireworks proxy. The final answers are validated, matched to their original task IDs, and written as valid JSON to /output/results.json. The V17.5 system achieved 94.7% accuracy, equivalent to 18 out of 19 tasks, while using 4,047 Fireworks tokens.