TL;DR
Built a multilingual email classification system using unsupervised learning and fine-tuned BERT. Handled acronyms, multiple languages, and noisy data. Achieved high precision in routing emails to the right recipients.
An organization was overwhelmed with inbound emails covering multiple departments, languages, and use cases. These messages arrived unstructured and were often full of acronyms and domain-specific terminology—making manual routing slow, error-prone, and unscalable.
I was tasked with building an automated system to classify and redirect incoming emails to the correct recipients—across teams, topics, and languages. The goal was to ensure accurate triage with minimal human intervention, despite noisy, inconsistent input.
To uncover the latent structure of the email data, I first applied unsupervised learning techniques like TF-IDF and k-means clustering to explore the types and number of classes. I validated the resulting clusters in close collaboration with domain experts to ground the taxonomy in real-world needs. Once the structure was established, I fine-tuned a multilingual BERT model, optimized for precision across multiple languages and specialized terminology. I developed data synthesis and preprocessing pipelines to clean inputs and handle noisy edge cases. The model was trained to recognize acronyms, abbreviations, and jargon—critical in a multi-domain setting.
The final system achieved high precision in email classification, significantly reducing manual routing errors and improving response times. It effectively handled multilingual content and complex terminology, and was deployed into production to streamline communication workflows across the organization.