Commercial Relevance Classifier

Developed a commercial relevance classifier that processed over 200k+ user searches to enhance a recommendation system, effectively matching searches with relevant advertisers with an impressive 92% accuracy. This initiative was pivotal in driving machine learning efforts that reduced cost per action by 10%, showcasing the impact of AI-driven insights in a fast-paced startup environment. The classifier played a crucial role in automating decision-making processes, allowing for faster, data-driven responses to evolving market needs and customer preferences.

The project involved designing, building, and fine-tuning open-source LLM models, including LLaMA 3.1, which were hosted on AWS to tackle challenges in ad copy generation. These models were meticulously optimized to deliver contextually accurate outputs, and the work was presented at AWS’s AI accelerator program, directly contributing to the startup’s successful fundraising and client acquisition. The demonstration highlighted the potential of advanced NLP techniques in real-world applications, drawing significant interest from investors and partners.

Additionally, the project explored advanced Retrieval-Augmented Generation (RAG) optimization methods, such as Self-Extend from ICML 2024 and Compressing Context from EMNLP 2023, to address long-context challenges commonly faced in large language models. By comparing various embeddings available on Hugging Face, the project identified cost-efficient solutions that balanced performance with scalability. Continuous improvements were facilitated through Amazon SageMaker, leveraging automated retraining pipelines to keep models up-to-date with the latest data, ensuring that the system remained robust, adaptive, and highly performant in a dynamic environment.