TokenSmart Router is a ZERO_API_CALLS hybrid agent for Track 1 of the AMD Developer Hackathon: ACT II. It handles all eight capability categories through a three-layer cost ladder, using the Fireworks proxy only as a last-resort fallback that never triggered on our practice set. The insight: Track 1 rewards fewest Fireworks tokens above an accuracy gate, and local computation counts zero. So instead of routing to a cheap Fireworks model, we asked how many tasks we can answer without calling Fireworks at all. On the practice set, one hundred percent. Layer 0 is a regex classifier that buckets each task into one of eight categories. Layer 1 is deterministic Python solvers: SymPy for math, python-constraint for CSP logic puzzles, regex for NER, confidence-gated keyword scoring for sentiment, and a curated lookup for factual questions. Layer 2 is Qwen2.5-1.5B-Instruct in 4-bit GGUF, bundled in the Docker image and running on CPU via llama-cpp-python — handles summarization, code generation, and debugging with tight prompts and hard token caps. Layer 3 hits Fireworks only if both layers return empty. Engineered for 4 GB RAM, 2 vCPU, CPU-only. Qwen 1.5B at Q4_K_M occupies ~1 GB. llama-cpp-python compiled with GGML_NATIVE=OFF. Final image 3.15 GB, cold start ~5 sec. Result: 5/5 practice tasks answered with zero Fireworks tokens — qualifies for the ZERO_API_CALLS flag per the participant guide. Stack: Python 3.11, Docker linux/amd64, Qwen2.5-1.5B, llama.cpp, SymPy, python-constraint, Fireworks AI.
Category tags: