Given the engineering complexity of NLP and machine learning architectures, what approach do you and your team use for implementing deployable NLP systems? For instance, how do you decide when and what level to use open-source tools versus building your own custom tools?