Our custom Docker registry and containerd adapter that loads ML models up to 15x faster – cutting down cold start times by up to 90%.