Machine Yearning

This project implements an ultra-efficient, production-ready hybrid LLM orchestrator designed specifically to maximize accuracy while minimizing token consumption on resource-constrained grading infrastructure. At its core sits a deterministic this routing engine analyzes incoming prompts using hyper-calibrated keyword heuristics. Simple, general-knowledge queries are processed entirely offline via a localized Llama-3.2-3B model, delivering instant answers at zero token cost. Conversely, complex "trapdoor" tasks—such as Python code debugging, named entity recognition (NER), and advanced logic puzzles—are instantly identified and routed to the high-performance cloud API fallback, kimi-k2p7-code. The specialized, highly restrictive system prompts isolate the cloud model’s execution, suppressing chatty conversational preambles and forcing immediate, auto-grader-compliant outputs. Operating with multi-threaded configurations tailored to the evaluation server's 2 vCPUs and strict 4GB RAM threshold, this decoupled architecture provides an elite, predictable, and fully scalable solution that guarantees accuracy on the very first attempt.