HALCYON·AI

Halcyon AI ships a new model every week without a single rollback.

52: model releases a year
0: customer-facing rollbacks
5%: first-step exposure
9: judges per release

The challenge

Halcyon’s document model improved every week in the lab, but each release to customers was a quarterly event with a war room and a rollback plan.

The approach

Each new model now starts at five percent of traffic. Nine judges score its answers against the current model, and an experiment decides when it moves up.

The result

Halcyon ships a model every week. In twelve months no customer has seen a model that failed its judges.

“Every model gets five percent of traffic and a jury of judges. The good ones graduate on their own.”

Aiko Mori Head of ML, Halcyon AI