Token-Efficient Router

Our agent handles 19 hackathon tasks in parallel using an 8-worker pipeline. Each task first hits deterministic solvers — a 659-entry knowledge base for factual questions, sympy-powered arithmetic with word-number conversion, word-set sentiment scoring with negation handling and contrast detection, regex NER covering 235+ entities, 50 pre-written code templates for common programming tasks, and AST-based debugging that catches syntax errors, mutable defaults, and off-by-one logic. Tasks the local tier can't resolve (novel code, unfamiliar entities, complex logic) fall back to Fireworks AI cloud through a three-retry chain. We route by category — deepseek for math, kimi for code, gpt-oss for factual — and drop to the next model only on failure, with Claude as the final resort. The container is 345MB, runs on CPU alone, and clocks each task at 30 seconds max.