.png&w=256&q=75)
🤓 Latest Submissions
.png&w=828&q=75)
Agent Rerouter
ZeroToken Agent is a general-purpose AI routing system built for the AMD Developer Hackathon Track 1 challenge, where the objective is to maximize answer accuracy while minimizing Fireworks API token usage. Instead of relying on cloud inference for every request, our system is designed around a layered execution strategy that performs as much computation locally as possible. The agent first classifies incoming tasks into categories such as factual question answering, mathematical reasoning, summarization, sentiment analysis, named entity recognition, logical reasoning, code debugging, and code generation. Whenever a problem can be solved deterministically, the router bypasses language model inference entirely. Mathematical templates are solved using rule-based arithmetic, while named entity extraction combines lightweight NLP with pattern matching to produce accurate structured outputs without consuming any cloud tokens. For tasks requiring natural language understanding, the system uses a bundled 4-bit quantized local language model running entirely inside the Docker container. Because inference occurs locally, these computations contribute to answer accuracy while consuming zero Fireworks API tokens under the competition rules. The model is loaded once during container startup and reused throughout execution to minimize latency and memory overhead. The entire solution is engineered for the official evaluation environment of 2 vCPUs and 4 GB RAM. Thread counts, context size, memory usage, and startup behavior were optimized to remain within resource limits while maintaining reliable performance. The container always generates a valid output file and exits successfully, even under failure conditions, ensuring robust evaluation. By combining deterministic computation, efficient local inference,and resource-aware engineering,ZeroToken Agent demonstrates that high-quality AI task execution can be achieved with minimal cloud dependency and competitive token efficiency
13 Jul 2026
