AstraRoute AI Agent #10

This agent is built for the AMD Developer Hackathon (ACT II) Track 1. It reads an array of tasks from a JSON input file, intelligently routes each task through a multi-stage pipeline, and writes answers back to a JSON output file. The core strategy is token minimization without sacrificing accuracy: A rule-based task classifier categorizes every prompt into 8 types: code, math, sentiment, ner, summary, factual, reasoning, and general Deterministic local solvers handle math expressions and clear-cut sentiment — costing zero Fireworks tokens A smart model router selects the most cost-effective allowed model per task type (e.g., Kimi for code, MiniMax for factual) A 404-aware fallback chain automatically retries with the next allowed model if one is unavailable The entire batch runs in a single Docker container (linux/amd64), exits cleanly with code 0, and never hardcodes API keys or model IDs The agent is fully tested (55 unit tests), containerized, and designed to run within the grading harness constraints: 4GB RAM, 2 vCPU, ≤10 min total runtime.