Shipping AI should be as safe as shipping code.
Repath is the deployment layer for AI. It routes a slice of real traffic to a new prompt or model, judges every answer, and rolls back on its own when quality drops — before most of your users ever see the change.
Software teams stopped shipping changes to everyone at once years ago. Canary releases, feature flags and automatic rollback are standard for code.
AI changes still go out all at once. A reworded prompt or a new model version reaches every user the moment it is deployed, and when it is worse, nothing errors — the answers are just wrong, politely. Teams find out from support tickets.
It brings the same discipline to prompts and models. One line points your existing OpenAI-compatible client at Repath. From then on, every change can go out as a canary: a few percent of traffic first, a quality gate at each step, and a rollback the moment the judge says the new version is worse.
Every decision comes with its evidence — the answers that moved the score, and the judge's reasons.
Out of your way
The gateway sits in your request path, so its own overhead is about 0.4ms at the median — and if it ever cannot serve, it says so rather than degrading your traffic.
Evidence over vibes
A rollout advances on judged quality from real traffic, and never on a health check alone. Every decision is logged with the numbers that made it.
Where your data must be
Use the cloud, or, when data must not leave your network, we run the whole stack inside it.