Hybrid Token Efficient Routing Agent

This project is a routing agent that decides, per task, whether a small local model can answer correctly or whether the query needs to be escalated to a larger model hosted on Fireworks AI. The core idea: a local Qwen model, run inside the container, attempts every task first. Alongside its answer, it self-reports a confidence score. If that score clears a threshold, the local answer is used and the task costs zero tokens toward the hackathon's scoring, since local inference isn't billed. If confidence falls below the threshold, the task is escalated to a Fireworks-hosted model via FIREWORKS_BASE_URL, using only models from the harness-provided ALLOWED_MODELS list. The agent is built to handle all eight task categories specified for Track 1: factual knowledge, mathematical reasoning, sentiment classification, text summarization, named entity recognition, code debugging, logical/deductive reasoning, and code generation. It reads tasks from /input/tasks.json and writes results to /output/results.json in the required schema, and reads all Fireworks credentials and model IDs from environment variables at runtime rather than hardcoding them, per the harness rules.