At Typeform, our customer support agents handle hundreds of support chats and several thousand support tickets each month to help with all the problems our customers have, so we decided to use this direct communication with our customers to deliver the best possible service and value.
In January 2022, we launched Actions After Support, where the overall goal is to understand the relative value of different actions by our customers and identify the best next action for them based on the customer profile and the stage of their customer journey.
We created a recommender model based on customer similarity that is served using a custom UI to our support agents, so they can use this information while having a live chat with our customers.
The model has been written in Python, and it is served using an Airflow and ML Gateway, to a custom UI created on streamlit, allocated as a docker container on a Kubernetes pod. The agents can also send direct feedback about the performance of the model to the data team.
In this talk, I will share with you the whole pipeline used from building the recommender model, serving the results, showing them on a custom UI and the monitoring that is in place, as well as the collaboration between stakeholders and data team. A complete and successful data science use case.
