Using few-shot learning, I trained a multi-label machine learning classifier to classify emails into topics from the text of the email. As there was no existing labelled data, I used SetFit to fine tune a Sentence Transformer using contrastive learning. The model was deployed in 8 weeks to an AWS SageMaker endpoint which is serving 50k requests per minute in production. The deployed model was able to detect malicious “conversation starter” phishing emails which were previously undetected with a precision of 90%, and reduced the level of spam by ~20%. This allowed the company to demonstrate detection improvements to potential clients as part of sales PoVs.

Led the research and deployment of the first full end to end machine learning phishing email detection algorithm at Tessian. Trained a CatBoost machine learning classifier based on features derived from email metadata, and developed an associated feature engineering pipeline based on PySpark. Devised a method to estimate the precision of the system on production data, which reduced time to deployment. The model detected 20% more phishing emails compared to the existing system. This enabled the company to demonstrate to potential clients improvements in the efficacy of the system when compared to competitors.

Developed a prototype semantic search system using dense text passage embeddings, ElasticSearch and zero-shot classification. This allowed users to search complex government climate policy PDF documents to understand important climate change themes. Led the team which delivered a highly functional prototype with a NextJS front end to time despite a stretching target to deliver the system in 8 weeks. It enabled the company to demonstrate their capability at the COP 26 international climate conference and secure partnerships with potential funders and organisations.

Responsible for the research and development of a machine learning system to automate the classification of attributes from product descriptions sourced from retailers. After demonstrating the potential of the system to senior leaders and securing funding for development, I delivered a scalable system using FastText which was able to automate the classification of thousands of product attributes with high accuracy. It was estimated to deliver ~$2M in cost savings when compared to manual data entry which resulted in the system being deployed and integrated as part of the company’s global data entry system.

Built a text data linkage system to match product descriptions obtained from consumer receipts to descriptions received from retailers. The system needed to be delivered quickly to remove a significant roadblock which would have prevented a major new service being made available to clients. I developed a baseline model using multiple matching features coupled with a Logistic Regression model. As part of this, I employed cross validation to overcome issues with domain shift when the model was applied to new retailers. The baseline model was developed in 4 weeks to demonstrate that the problem could be solved and secure additional funding for development of the full solution.