.png&w=256&q=75)
🤓 Latest Submissions
AsterRoute: Zero Token Binary LLM Router
AsterRoute is a hybrid, token efficient general purpose agent built for Track 1 of the AMD Developer Hackathon: ACT II. Instead of spending LLM tokens to decide which model should answer, it runs a fine tuned 66M parameter DistilBERT classifier locally. The router predicts whether the inexpensive Fireworks tier can answer reliably and escalates uncertain requests directly to the strongest allowed tier. The system combines learned confidence thresholds with auditable category rules. Factual knowledge, math, logic, sentiment, code debugging, and code generation default to the inexpensive tier, while named-entity recognition and summarization use a P(cheap_ok) confidence gate. Training uses 176 empirically labeled examples across eight categories, negative oversampling, class weighting, and an independent grading pipeline. Code-generation answers are executed against tests, while open ended outputs use structured judging. All answer generation remains on Fireworks. The local model only selects the answer model, so routing consumes zero API tokens. Offline replay achieved 98.9% accuracy, matching the strongest model baseline while routing 173 of 176 tasks to tier0. The Linux/AMD64 Docker image implements the required /input/tasks.json to /output/results.json contract and completed our end to end Fireworks test in approximately nine seconds. The project also includes threshold evaluation, a 72 prompt no overlap holdout suite, reproducible training scripts, automated smoke tests, and an interactive Streamlit demonstration. The training code runs without modification on CPU, CUDA, Apple MPS, or ROCm backed AMD hardware.
13 Jul 2026
