V3 Zero Token Hybrid AI Router

This project introduces a highly resilient, cost-optimized AI routing architecture designed to maximize token efficiency without sacrificing accuracy. Built entirely in asynchronous Python to ensure high-throughput execution within strict 10-minute timeout constraints, the system operates as a smart gateway between local edge computing and the Fireworks API. To aggressively minimize API token usage, the agent employs a deterministic semantic router that intercepts incoming natural language tasks. Lightweight, predictable tasks—such as Named Entity Recognition (NER), sentiment classification, and text summarization—are processed entirely on-device using a zero-token local stack featuring spaCy and a 4-bit quantized Llama-3.2-1B-Instruct model, engineered to operate safely within the strict 4GB RAM ceiling. Conversely, complex workloads demanding high-tier reasoning, logic, and code generation are dynamically routed to the Fireworks API. The system parses the allowed models at runtime, intelligently mapping coding tasks to specialized models like kimi-k2p7-code and mathematical puzzles to heavy reasoning engines like gemma-4-31b-it. By enforcing strict temperature controls and ultra-compact system prompts, this hybrid architecture delivers an elite accuracy-to-token ratio, serving as a blueprint for enterprise deployments seeking scalable, cost-effective generative AI solutions.