I led the technical design and implementation of a BERT based multi-label multi-class classification model that is used to this day to classify all types of customer feedback according to an internal multi-level taxonomy. The work included tweaking Google’s BERT model to support multi-label classification, optimizing the code, building the infrastructure and serving layer and deploying it at scale.
The system currently powers high-impact use cases that rely on the automated classification of customer feedback including product reviews, return comments and customer questions at scale, serving a daily throughput of 500k+ texts per day. For this project, I collaborated closely with world-class scientists and researchers, translating their state-of-the-art models into production grade customer-facing services.