We have removed an upstream from our API load balancer that seems to be accounting for all of the timeout errors and we are seeing an immediate improvement.

We are investigating a network degradation between two data centers that is affecting a subset of predictions.