MobZ - Smart Router

MobZ is a hybrid token-efficient AI runtime designed to solve diverse natural language tasks while minimizing inference cost without sacrificing accuracy. Instead of sending every request directly to a large language model, MobZ introduces a lightweight cognitive pipeline that analyzes each prompt before any Fireworks API call is made. The runtime begins by loading incoming tasks from the evaluation JSON and applying a semantic prompt compression stage. Unlike traditional summarization, this component rewrites prompts into a more token-efficient representation while preserving their original intent, constraints, output format, entities, numerical values, and programming context. The compressed prompt is then processed by a custom-trained Cognitive Analyzer based on ModernBERT. Rather than generating text, this encoder performs a single forward pass to predict multiple characteristics simultaneously, including task category, estimated difficulty, reasoning depth, expected output format, and expected response length. This design provides deterministic, extremely fast inference while consuming minimal computational resources. Using this cognitive profile, MobZ consults an embedded benchmark database generated through MobZ Bench. This database stores benchmark-derived performance information for the Fireworks models available during evaluation, allowing the runtime to estimate which model offers the best trade-off between accuracy and token cost for the detected task and complexity. The routing engine then dynamically selects the most efficient allowed Fireworks model and forwards the request through the official Fireworks API. Every decision is entirely benchmark-driven and adapts automatically to the list of models provided at runtime. MobZ is fully containerized, requires no external orchestration services, and complies with the competition requirements by reading tasks from the provided JSON input, routing all remote inference through Fireworks.