Adaptive Model Dispatcher

Adaptive Model Dispatcher (AMD) is a highly optimized routing agent built for Track 1 of the AMD Developer Hackathon: Act II. Because Track 1 ranks submissions by token count after passing a strict Accuracy Gate, any single-model approach is either too expensive or fails specialized tasks like math or coding. AMD solves this with a 5-layer sequential dispatch pipeline: 1. Prompt Compression: Strips unnecessary whitespace and collapses blank lines, saving 10-15% of input tokens on every API call. 2. Tier-0 Deterministic Solver: Safely parses and evaluates basic math, percentages, square roots, temperature, and unit conversions locally for 100% accuracy at 0 token cost. 3. Smart Triage: A regex-based heuristic classifier categories tasks instantly. Ambiguous cases fallback to a fast, cheap classification LLM. 4. Optimal Dispatch: Routes coding to Kimi K2.7 Code, general tasks to MiniMax M3, and sentiment/NER to a local Qwen 2.5 1.5B running inside the container (0 Fireworks tokens). 5. Speculative Validation: Catches refusals, formatting issues, or truncation, and triggers category-specific corrective retries. AMD also leverages the Gemma 4 Bonus Prize eligibility through a Speculative Gemma Try-First strategy coupled with a dynamic Circuit-Breaker to safeguard the 10-minute container runtime limit.