OyeHoye Zero Tokens

Zero Tokens (OyeHoye Router) is a local-first AI task router built for the AMD Developer Hackathon Track 1. It answers eight categories of natural-language tasks (factual, math, sentiment, summarization, NER, code debug, code gen, and logic) while spending zero Fireworks tokens at grade time. The design is cost ordering. Each task takes the cheapest tier that answers it correctly: deterministic Python solvers handle exact arithmetic, word problems, logic-grid puzzles, and common code functions; spaCy en_core_web_md handles named entity extraction; and a bundled Google Gemma 4 E2B model running locally via llama.cpp handles the rest. Fireworks is kept out of the runtime path entirely. The client exists for offline testing but is never called during a graded run and is not compiled into the image. Math uses program-aided reasoning: the model writes a short Python program, a restricted sandboxed subprocess executes it, and the output layer extracts the final answer. Chain-of-thought is off, tested and proven worse on this 2B model. Every deterministic tier abstains rather than emit a wrong answer, and abstained tasks drop to the next tier, so the results file is always complete. The runtime ships as compiled native extensions. Cython translates the Python source to C and gcc compiles it to .so, so the final Docker image carries no readable source, only the compiled modules, the spaCy model, llama.cpp, the Gemma weights, and a three-line launcher. The image is pinned to linux/amd64 to match the grading box. The entire pipeline runs on 2 vCPU with 4GB RAM in under 10 minutes. Peak observed memory is 3.66 GiB, a 19-task local run completes in 34 seconds, and a 64-task stress run in 111 seconds. Built by Nikhil Kapila, Gaurav Gosain, and Zain Raisan. Sponsored by Google Gemma.