.png&w=256&q=75)
🤓 Latest Submissions
hybrid-routing-agent
hybrid-routing-agent is a general-purpose AI agent built for AMD Developer Hackathon Track 1, designed to handle all eight required capability categories — factual knowledge, mathematical reasoning, sentiment classification, text summarization, named entity recognition, code debugging, logical reasoning, and code generation — while minimizing token cost. Rather than sending every query to a large remote model, it scores each incoming prompt on a 0–1 complexity scale using signals like query length, code/math keywords, multi-step reasoning cues, and format-sensitive tasks like JSON extraction. Simple queries are answered entirely by a free local model (phi3:mini via Ollama, bundled directly into the Docker image), harder queries cascade to Fireworks AI only when the local model's confidence is low, and clearly complex queries route straight to Fireworks. This keeps the accuracy gate satisfied while keeping the total token count — the metric the leaderboard ranks on — as low as possible, since local inference costs nothing and only Fireworks-routed tokens are scored. The submission packages this as a self-contained Docker image implementing the competition's batch contract: it reads a JSON array of tasks from /input/tasks.json, processes them concurrently with a deadline-safe thread pool, and writes results to /output/results.json, reading the harness-injected API key, base URL, and allowed model list purely from the environment at evaluation time.
13 Jul 2026
