Noob_Router

Our project is a hybrid token-efficient AI agent designed to maximize accuracy while minimizing inference cost. Instead of sending every request to a large cloud model, the system intelligently routes each task through a multi-stage decision pipeline. Every request is first processed by a lightweight local LFM model running entirely inside the container, resulting in zero cloud token usage. The response is then evaluated using task-specific validation, confidence estimation, and self-consistency checks. If the local answer is considered reliable, it is returned immediately. Only low-confidence or complex tasks are escalated to the Fireworks AI API, where the router dynamically selects the most cost-effective model from the allowed runtime models. The agent supports a diverse set of capabilities, including factual question answering, mathematical reasoning, sentiment classification, text summarization, named entity recognition, code debugging, logical reasoning, and code generation. Each category uses specialized prompts and validation rules to improve accuracy while keeping responses concise. The system also minimizes token consumption through prompt compression, category-specific generation limits, and adaptive routing decisions.