325 Agent — Hybrid Token-Efficient Router

325 Track 1 Agent is a hybrid token-efficient router for AMD hardware. 11 deterministic local solvers handle 93.2% of tasks at zero tokens on AMD EPYC — no API, no GPU, sub-millisecond. Complex tasks route to Fireworks AI on AMD Instinct MI300X via ROCm 6.0 with tiered token allocation and dual-model consensus. Key metrics: 69/74 local (93.2%), 96.8% savings (600 vs 18,500 tokens), 54.5MB Docker, 125 test tasks, 50 adversarial edge cases, published failure taxonomy with zero silent failures. Solvers: math, percentages, unit conversion, temperature, sentiment, counting, string ops, logic, knowledge facts (70+), named entity extraction, regex validation. Architecture: Task Classifier → Local Solvers (0 tokens, <1ms) OR Fireworks AI (tiered 100/500/1000) → Dual Consensus on complex prompts. Simulation mode works without API key. Full per-task routing transparency. Business case: $1.2B→$8.7B API gateway market. Cost savings from $5K/yr (solo dev) to $5.3M/yr (platform scale). Revenue model: MIT open source → managed API. Built solo in under a week. 54.5MB alpine Docker. GitHub CI. MIT license. Compared head-to-head against Token Frontier Model — 41x more efficient on 4x more tasks.