Customer Obsession Ticket Assistant (COTA) is an applied Machine Learning and Natural Language Processing project I worked on my first year at Uber AI Labs. It was a collaboration between Ai Labs, Applied Machine Learning, Michelangelo (Uber's machine learning infrastructure team) and the customer obsession team.
The insight was to provide the Customer Support Representatives at Uber with a tool to help them making better and faster decisions. They usually have to select among thousands of possible ways to classify and resolve a ticket, which is time consuming and error prone. So we built models to suggest them the most likely options given the request from the user and some contextual information regarding the trip their complaint is about.
I worked on the v2 model using Deep Learning architecture for multi-task learning. The model was capable of reducing the handling time of the ticket by more than 16% without any degradation in customer satisfaction.
We published an article at KDD describing the models and two articles on the Uber Engineering Blog, the first about the problem and the models, the second about the deployment and scaling (in which I had just a little involvement). I presented the results of this work both at scientific conferences (KDD 2018 in London) and industrial ones (O'Reilly Artificial Intelligence 2018 San Francisco conference).
Our effort was appreciated both by a Comet ML article on Medium and an article form Algoritmia that defined it "One of the leaders in the space of Machine Learning for Customer Support". Harvard Business Review features it in an article about "Machine Learning: The Next Generation of Customer Experience".