ProsoponRouter

CascadeSolo Router (ProsoponRouter) is a hybrid, token-efficient AI agent built for Track 1 of the AMD Developer Hackathon. Instead of routing every task to a large language model, it first classifies each incoming prompt and attempts a deterministic, zero-token solution wherever one reliably exists. Math problems are solved with safe arithmetic evaluation, sentiment analysis uses VADER, and named entity recognition uses spaCy — all without spending a single token. For harder categories — factual knowledge, summarization, logical reasoning, code debugging, and code generation — the agent escalates to Fireworks-hosted models (MiniMax M3 and Kimi), with token budgets carefully tuned per category to balance accuracy against cost. An optional local Gemma model runs on-device as a further zero-token tier, validated with the same strict correctness checks — including a custom function-signature check that ensures generated code preserves the exact requested function name, and a reasoning-scaffold guard that rejects incomplete or meta-commentary answers before they're accepted. The full pipeline runs inside a single Docker container meeting the harness's strict constraints: 2 vCPU, 4GB RAM, a 10-minute timeout, and linux/amd64 compatibility, reading tasks from /input/tasks.json and writing results to /output/results.json