Replacing a Live AI Pipeline Without Downtime

This is an AI assistant product on web and mobile, across 5 repositories. We owned the architecture. We wrote the specifications that other engineers built from.

We added a new system that runs AI agents beside the live AI pipeline, instead of a rewrite. Switches choose the old pipeline or the new path. The new path has its own database.

We removed the canary and the percentage rollout. Both send a change to some users first. We replaced them with one switch. The new path stays off unless a second safety switch, the grounding gate, is also on.