GemmaGate

GemmaGate is a local-first AI routing agent built for the Hybrid Token-Efficient Routing Agent track. It handles eight task types: factual knowledge, math, sentiment, summarization, named entity recognition, code debugging, logical reasoning, and code generation. Instead of sending every prompt directly to a large model, GemmaGate first classifies the task and tries deterministic local solvers for cases that can be safely verified, such as arithmetic, sentiment labels, entity extraction, logic puzzles, constrained summaries, and common code patterns. Each local answer is checked by a validator before it is accepted. If the answer cannot be verified locally, GemmaGate routes the task to the cheapest available Fireworks model and escalates only when validation fails. The goal is simple: preserve accuracy while spending as few Fireworks tokens as possible. In local benchmark testing, GemmaGate solved 21 of 23 tasks fully locally and achieved 23/23 accuracy with only 2 Fireworks calls.