HybridRoute

Created by team LLM Learner on July 06, 2026
Hybrid Token-Efficient Routing Agent

HybridRoute is a hybrid token-efficient routing agent built for the AMD Developer Hackathon Track 1. It classifies each prompt with an estimated workload score first, then decides whether to answer locally or call a remote model using an API key. Simple tasks (with an estimated score of 0–35) stay on a local Ollama model for low latency and zero Fireworks tokens. Mid-range tasks try local first with a verifier and escalate only when confidence is weak. Hard or high-risk tasks (score 61–100) go to a remote model, with prompt compression to keep token use down. The judged submission is a public Docker image that reads /input/tasks.json, routes each task to an allowed remote model on Fireworks, and writes /output/results.json. Alongside it, a React demo UI shows the same routing story live: Local vs Hosted, model name, runtime, and remote tokens under each reply in one chat thread. The goal is simple: spend tokens only when quality requires them, without sacrificing accuracy on difficult work. HybridRoute can do it all.

Category tags: