TokenPilot: Adaptive Fireworks Model Router

TokenPilot is a professional, containerized AI routing system designed for efficient model selection. It analyzes each incoming task and determines whether it can be handled locally or should be sent to an appropriate Fireworks model. This approach reduces unnecessary inference costs while maintaining reliable answer quality. The system supports structured batch processing through /input/tasks.json and produces standardized results in /output/results.json. It uses a local routing layer, confidence-based escalation, dynamic model selection from the evaluator-provided model list, persistent answer caching, prompt-cache affinity, retry handling, and response validation. TokenPilot is designed to handle a broad range of tasks, including factual questions, mathematical reasoning, logical deduction, summarization, sentiment classification, named entity recognition, code debugging, and code generation. The application is packaged as a portable linux/amd64 Docker container and uses runtime-provided Fireworks credentials and endpoints without storing secrets in the project. The goal is simple: deliver accurate answers while using the smallest practical amount of paid model inference.