Reduce inference time for BERT models using neural architecture search and SageMaker Automated Model Tuning