Developed a Transformer based workflow on behalf of Employee Relations to classify employee incident descriptions and identify potential code of conduct violation cases.

• ~ 2 million USD saved

  This project helped the Employee Relations Team save countless hours of manual inspection and reduce costs roughly by 2 million USD.

• A fully deployed workflow

  Collaborating with the Data Engineering Team, we developed and deployed an end-to-end workflow. This workflow processes comments from past surveys, applies our ML model, and alerts Employee Relations if a survey response indicates a code of conduct violation based on predicted probabilities.

• 30% Increase in Recall

  By leveraging a Sentence Transformer to generate contextualized embeddings and conducting extensive Grid Search hyperparameter tuning experiments, we achieved an impressive 30% increase in recall for the positive class (Code of Conduct Violations)

• Custom Metrics for Automation

  Introduced ARR (Percentage of comments that will bypass manual review) and ARA (Accuracy of target labels for bypassed comments) in order to measure and minimize the percentage of comments that would need manual inspection. Achieved ~99% for both ARR and ARA.