By 2020, every Samsung product will be AI enabled, creating entirely new experiences and value for users. In this talk, we describe how the fundamental technologies such as automatic speech recognition and natural language understanding are being used to create these new experiences. Bixby is Samsung's AI Platform the core of which is Bixby Voice – an intelligent, personalized Assistant that enables users to use Voice for Samsung or 3rd party services. Bixby learns, evolves, and adapts to you to help you get things done seamlessly. Each feature of Bixby is designed to make your life easier, whether it's through touch, type, or voice. We talk about the evolution of Bixby Voice & associated research & development work at Samsung (including R&D centers in India). We talk about challenges in domain & intent classification for Natural Language understanding, and Deep Learning based solutions ranging from CNN vs. RNNs, Word vs. Character based models, Handling of variations of data and conflicts between data. We also describe recent efforts and results in improving speech recognition (ASR) performance for end users of Samsung's Bixby voice assistant. One of the main challenges with ASR is speaker variability. Different users, depending on their style, accent and other traits can pronounce the same word or sentence in many different ways. Hence, a one-size-fits-all approach works poorly for some segment of users, even though aggregate ASR performance may be quite high. Specifically, we present two techniques of improving ASR performance by considering the end user. The first is implemented at a macro level and uses a Hierarchical Accent Determination system dealing with accent variations. For English, eight accents [GB, US, Australian, Canadian, Spanish, Korean, and Indian & Chinese] are identified at macro level and accent-specific models corresponding to the identified accents are used. Similar approaches are applicable for languages like Chinese, Hindi, Arabic, which have significant accent variations at regional level. The second technique involves adapting the ASR models for a particular end user. Here, we describe how the acoustic model can be enhanced for a particular user's speaking style by training a new model on-device for that user. The key challenges in this problem are how to select utterances for training and adaptation, how to generate ground truth labels and handle errors in ground truth estimation.