TOKENMAN is a hybrid, token-efficient routing agent for general-purpose AI tasks. It reads tasks from /input/tasks.json, identifies the capability each prompt requires, and chooses the lowest-cost reliable path. Straightforward math, entity extraction, sentiment checks, summaries, logic patterns, and narrow code fixes use deterministic Python tools when possible. More complex factual, reasoning, and code-generation requests are routed to approved Fireworks models with compact prompts, tailored model selection, output validation, and selective retry. TOKENMAN writes valid JSON results to /output/results.json. It is built to maintain strong accuracy across diverse tasks while using as few model tokens as possible.
Category tags: