Recommendation- Created a recommendation model for newly released shows using SBERT, which outperformed the Byteplus model by 5% based on 7-day averages.- Developed the ALS recommendation model, which became the foundation for most sections in the KUKU FM app. Although we later transitioned to more sophisticated models, this remains the foundational model for most sections in our app.- Through these improvements and comprehensive data analysis, we successfully replaced Byteplus, leading to $20,000 in monthly cost savings.- Currently implementing a two-tower recommendation model for new users, aiming to increase average listening minutes during their first week.Predictive Models- Created a churn prediction model for subscriptions that identifies at-risk users with 95% accuracy, allowing the team to implement targeted retention strategies.- Developed a churn prediction model for the coin business and conducted experiments with likely-to-churn users, which helped improve both revenue and listening engagement.- Led a project predicting ad engagement using SBERT embeddings for scripts and vision models for video content to forecast user pause rates while scrolling through ads. LLM projects- Created an internal tool that helps the business and product ask questions to Kuku FM data through Slack in nautual language. This tool uses NL2SQL technology and was built using Langchain and OpenAI.- Implemented OpenAI to generate dynamic tags from show scriptsOther Projects- Implemented a Learning-to-Rank (LTR) model for coin pack rankings, improving store conversion rates and revenue by optimizing purchase flows.- Created an anomaly detection model using the isolation forest algorithm