Developed a Netflix-like real-time recommendation system for over 20 business units in Asia and Europe, increasing sales by 10 HKD per user and average online time by 4.6 minutes. Created customer and item embeddings using the LightFM model, achieving MAP@6 of 6%. Optimized a re-ranking algorithm on Apache Spark, reducing runtime from 5 hours to 20 minutes. Developed a churn prediction model and customer segmentation using Kafka for real-time predictions with 80% precision and 40% recall. Managed the feature store platform, ensuring data quality and pipeline integrity, and reducing duplicate efforts across teams. Optimized ELT pipelines using Delta Tables and SCD Type 2. Implemented MLOps and Data Pipeline Ops for model management and deployment.

Developed a real-time e-commerce recommendation system for Vnshop, increasing the shopping rate by 147% for returning customers and 28% for new customers. Analyzed the taxi ecosystem using hotspot forecasting and churn prediction to boost revenue for VNPAY’s ecosystem. Created algorithms for VNTicket to maximize revenue through cross-selling and improving customer loyalty. Built a fraud detection system for QR transactions using Association Rule Mining, saving up to $150k USD in marketing campaigns with 87% accuracy for users and 79% accuracy for transactions. Managed and maintained the development environment for the Data Science team, including Spark clusters, Conda, Neo4j, and ETL templates.

Developed an end-of-month revenue forecasting model based on daily product values, achieving 80% accuracy. Conducted A/B testing and utilized a multi-arm bandit approach to optimize the homepage, leading to a 66% improvement in the Click-Through Rate (CTR) for new users. Developed an app recommendation system for millions of users, reducing time to action by 30% compared to previous models. Conducted social network analysis to identify patterns in user relationships and provided audience segments for marketing teams.

Drop Off Recommendation: Developed a large-scale recommendation system for trip destinations, handling approximately 50,000 trips per day, increasing the booking rate by 12% and reducing average lookup time by 20 seconds. Multi Dispatch: Implemented a real-time dispatch system for riders and drivers, achieving a maximum latency of 5 seconds. Be Map Analysis: Conducted analysis on Be Map effectiveness, focusing on driver and rider behaviors and anomaly detection using Gaussian Distribution Analysis (GDA). Applied low-pass filter and Kalman filter to remove noise from GPS signals, providing an overview comparison between internal Be Map search and Google Map search.