
TokenSmart Router is a ZERO_API_CALLS hybrid agent for Track 1 of the AMD Developer Hackathon: ACT II. It answers all eight capability categories and makes zero Fireworks API calls. The Fireworks fallback was removed entirely: zero tokens is the ranking mechanism, so one API call would drop us out of the zero-token tier, while one empty answer costs only a few percent of accuracy. Layer 0 is a regex classifier. Layer 1 is deterministic solvers, on the rule that a probabilistic system never decides anything a deterministic one can prove. SymPy handles math. python-constraint solves logic puzzles as CSPs. spaCy handles NER with a trained model, not regex. VADER handles sentiment with a 7,500-word scored lexicon that understands negation and intensifiers. A generated facts.json of 3,330 lookups is consulted before the model, so it is never asked a fact it might hallucinate. Layer 2 is Qwen2.5-1.5B-Instruct, LoRA fine-tuned on 5,657 examples on an AMD Instinct MI300X (loss 0.28), quantised to 4-bit GGUF (941 MB), running on CPU. Generated code is executed against a derived test case; failures are regenerated up to twice. Engineered for 4 GB RAM, 2 vCPU. Image 3.2 GB, ~110 s for 19 tasks. Result: 89-100% across ten independent test sets, at zero tokens every run.
13 Jul 2026