Tokomera: Adaptive Query Router

Tokomera is a general-purpose AI agent built for AMD Developer Hackathon ACT II, Track 1: Hybrid Token-Efficient Routing Agent. It is designed to answer natural language tasks across eight benchmark categories: factual knowledge, mathematical reasoning, sentiment classification, text summarization, named entity recognition, code debugging, logical reasoning, and code generation, while reducing unnecessary token usage on Fireworks AI The core idea is a local routing layer. Before calling a Fireworks model for the final answer, Tokomera uses a fine-tuned DistilBERT classifier inside the container to predict whether an incoming prompt is easy or hard. Easy prompts can be routed to a cheaper Fireworks model, while harder prompts can be escalated to a stronger model. Unlike prompt-based routing, which spends an additional LLM call just to decide where a query should go, Tokomera makes the routing decision locally. This means the routing step itself uses zero Fireworks tokens; only the final answer-generating model call is counted. The system supports multiple routing modes: always-cheap, always-expensive, prompt-based baseline routing, and fine-tuned local routing. This makes it possible to compare accuracy and token usage across strategies before choosing the most efficient submission configuration. The project is fully containerized and follows the Track 1 runtime contract: it reads tasks from /input/tasks.json, writes answers to /output/results.json, and uses runtime environment variables for Fireworks configuration and model selection. Tokomera’s goal is to pass the accuracy gate first, then reduce remote token cost through smarter local routing.