Hyper-Inference Router: AI Routing on MI300X

Most AI products pick one model for every request, regardless of what it actually needs — overpaying a frontier model to answer a greeting, or underpowering a request that needed real reasoning. Hyper-Inference Router classifies every incoming prompt and routes it to the model suited to the job: - Casual/simple queries -> a fast, cheap serverless model - Reasoning and creative tasks (proposals, strategy, marketing copy) -> Gemma 4 E4B (Google DeepMind), on a dedicated AMD Instinct MI300X deployment via Fireworks AI - Code-related queries -> a code-specialized model Classification is a zero-cost weighted keyword match — the router adds no LLM-call overhead of its own. This isn't a hackathon-only prototype. It's extracted from Kronos AI (kronos.devxhouse.com), a live SaaS platform generating job proposals, social content, and business documents for paying users today. This repo is a clean, standalone extraction — no Kronos business logic or proprietary code included. Every response shows real, live numbers: actual model used, latency, token count, cost in USD, and a live comparison against equivalent GPT-4o pricing — computed from that exact call, not asserted. Who this is for: any team running AI-generated content at scale that pays flat, per-call pricing regardless of task complexity. Kronos adopted this because a five-paragraph proposal and a one-line caption have different reasoning needs but were billed identically. This router plugs into any Fireworks-based product and cuts spend on requests that never needed a frontier model, without touching quality where it matters. Try it live: submit a prompt at the demo URL and watch it route in real time, with model, cost, and latency shown on screen.