Saleh Soltan is a Principal Applied Scientist at Amazon, where he has led the architecture design and pretraining strategy for models that became the Amazon Nova 1 and Nova 2 families, including key decisions around training recipes, architectural scaling, and training stability across multiple large-scale runs. He previously led pretraining for AlexaTM 20B, Amazon’s first company-wide generative model, enabling multilingual generation and synthetic NLU data creation. Earlier in his career at Amazon, he led BERT models pretraining and distillation efforts that transitioned Alexa from legacy NLU systems to neural architectures. He has also contributed to high-impact publications at venues including ACL, EMNLP, COLING, and SIGKDD, and holds a Ph.D. in Electrical Engineering from Columbia University.