Creating batch predictions through an inference pipeline