Bram Zijlstra is a Machine Learning Engineer at Media Distillery where he is responsible for the creation and development of software for TV Operators, Broadcasters and OTT Providers to improve the User Experience of their video services.
Passionate to apply scientific research and state-of-the-art technologies in new projects, Bram is known for his ability to rapidly build prototypes and turn them into complex software and his passion for Natural Language Processing, Computer Vision, and MLops.
Getting machine learning software to production is hard. Besides the usual complexity of lifecycle management, machine learning adds a whole new dimension of operations work. In order to make your architecture reproducible you now need to version your data, models, configurations, and new pipelines as well. Furthermore, the components in your ecosystem are more difficult to isolate, making debugging more difficult. Once all these components are put in place, there’s still no guarantee that your service works as intended, because by now chances are your data distribution or your concept of data has changed as well. This work is often either done by data scientists or software engineers, but both are not equipped to solve all these issues.
In this talk, we will go through some of the hard lessons we learned about putting machine learning into production. We will share some of the success stories, but also the failed projects and what we have learned from them.