TriForge: Hybrid Token-Efficient Router

TriForge is a production-grade Hybrid Token-Efficient Routing Agent that intelligently minimizes Fireworks AI token usage without compromising response accuracy. Instead of forwarding every request to a cloud LLM, TriForge autonomously analyzes each query and selects the most cost-effective execution path. The routing pipeline combines Semantic Intent Classification, Prompt Complexity Analysis, and Prompt Length Heuristics to determine whether a request should be processed locally or by Fireworks AI. Simple factual and conversational queries are answered locally using Ollama, consuming zero Fireworks tokens, while coding, mathematical, and reasoning-intensive tasks are routed to Fireworks AI only when necessary. To improve reliability, TriForge performs Local Self-Consistency Validation by generating multiple local responses and comparing them for agreement. If confidence is low or uncertain language is detected through Hedging Detection, the system automatically triggers Verify-Draft Escalation, where the remote model validates and refines the local draft instead of generating a completely new response. This significantly reduces unnecessary completion tokens while maintaining high-quality outputs. The framework also features Exact-Match & LRU Caching, a modular provider architecture, real-time analytics, benchmarking, and a containerized deployment-ready design. The analytics dashboard tracks routing decisions, latency, cache performance, token usage, and estimated cost savings, enabling continuous optimization. Built with Next.js, FastAPI, Ollama, Fireworks AI, SQLite, SQLAlchemy, and Docker, TriForge demonstrates how intelligent routing—not brute-force cloud inference—can build scalable, production-ready AI systems. By dynamically selecting the cheapest model capable of delivering accurate results, TriForge directly aligns with the challenge objective of maximizing accuracy while minimizing Fireworks token consumption.