Simply building a successful machine learning product is extremely challenging, and just as much effort is needed to turn that model into a customer-facing product. Drawing on their experience working on Zendesk’s article recommendation product, Wai Chee Yau and Jeffrey Theobald discuss design challenges and real-world problems you may encounter when building a machine learning product at scale.

Wai Chee and Jeffrey cover the evolution of the machine learning system, from individual models per customer (using Hadoop to aggregate the training data) to a universal deep learning model for all customers using TensorFlow, and outline some challenges they faced while building the infrastructure to serve TensorFlow models. They also explore the complexities of seamlessly upgrading to a new version of the model and detail the architecture that handles the constantly changing collection of articles that feed into the recommendation engine.

Topics include:

• Infrastructure for continuously changing textual data
• Deploying and serving TensorFlow models in production
• Real-world production problems when dealing with a machine learning model
• Data, customer feedback, and user experience