We started Floopy because our own AI bill scared us.

Floopy is the infrastructure layer between your agent and the providers — so the cheapest model that can do the job, does it. Every time. Without sacrificing what your users feel.

We were running AI features in production and the bill went from "line item" to "largest line item" in a few months. We tried what everyone tries: hand-written routing rules, token limits, aggressive caching, prompt compression. It worked, a little. It also broke constantly — the moment a new model shipped, our rules were stale. The moment a prompt changed, our caches were poisoned. The moment we relaxed anything, quality wobbled.

The pattern we wanted didn't exist as a product: route every call by a learned policy grounded in real user feedback, with canary-safe promotion and fast rollback. So we built it. Floopy is now an OpenAI-compatible AI gateway teams use to move traffic between models without moving their quality bar.

We believe three things, strongly: optimization belongs in the request path, not in a dashboard. Every routing decision should be explainable to a non-ML engineer on call at 3am. And teams should only pay for infrastructure that has measurably earned the money.