VoxRouter — Hybrid Token-Efficient Routing Agent

VoxRouter is a hybrid routing agent built for AMD Developer Hackathon ACT II Track 1: complete tasks using the fewest tokens possible while maintaining accuracy. Every incoming prompt passes through a 4-layer complexity classifier — rule-based pattern matching, keyword signals, structural analysis, and vocabulary entropy — that scores it from trivial to expert. Trivial and simple tasks are answered locally using quantized models (llama3.2:1b, qwen2.5:3b, phi3.5:3.8b) running on Ollama with AMD ROCm support, at zero API cost. Complex and expert tasks are routed to a remote model (Gemini 2.5 Flash Lite, with Fireworks AI as an alternate provider) for higher-quality generation. What sets VoxRouter apart from a standard router is confidence-based escalation: after a local model answers, VoxRouter evaluates its own confidence in that response. If confidence falls below threshold, the task is automatically retried on a remote model instead of returning a weak answer — so cost savings never come at the expense of correctness. The project includes a live React dashboard showing every routing decision in real time (route, complexity, tokens, cost, latency, confidence), streaming responses token-by-token, and a built-in 50-task benchmark suite spanning all five complexity tiers. Each benchmark run produces a single VoxRouter Score combining accuracy, routing correctness, and token savings — a reproducible way to prove the router actually works, not just anecdotally. The full stack is containerized with Docker Compose for one-command deployment: Ollama with ROCm GPU passthrough, a FastAPI backend, and the React dashboard.