Xidian Sun

2023-Current

• Tech led E-commerce Risk Control Security, including Traffic Security and Account Security for the US market.

Traffic Security

• Launched Ecommerce first anti-crawling and traffic protection system from 0 to 1, and scaled to all US.
• Designed and developed offline labeling algorithms and framework to identify over 50 million crawler requests daily across ~150 APIs.
• Trained and deployed real-time XGBoost-based detection models achieving >99.95% precision and 82% recall, continuously updated through automated daily pipelines.
• Developed session series features, trained and deployed LLM model to catch real device and simulator initiated bots, achieving 12% more recall.
• Trained and deployed real-time fraud detection models to distinguish between human and automated CAPTCHA solving behavior.

Account Security

• Designed and implemented risk level perception pipelines to detect account takeovers (ATO) across buyers, sellers, and content creators.
• Built an offline LLM-based text labeling algorithm to identify diverting messages on the platform, and using GPU for text embedding and calculating text similarity scores.
• Deployed a real-time BERT-based to intercept suspicious messages, blocking over 10k diverting messages per day with 100% precision based on Ops review.

2021-2022

• Designed and deployed tree-based model to predict trailer capacity utilization for bid optimization and trailer rebalancing decisions across regional markets.
• Improved forecast accuracy by reducing MAPE from 35% to 20% through iterative model tuning, temporal cross-validation, and feature importance analysis.
• Built trailer timelines projection to forecast trailer capacity used for bid decision and trailers rebalancing
• Developed robust ETL pipelines to extract and transform production-scale data for training, feature engineering, and model validation.

2019-2021

• End-to-end designed and developed bot detection system using tree-based model to protect Azure Frontdoor.
• Built statistical anomaly detection algorithms based on multi-dimensional traffic features to identify bot-like behaviors.
• Designed and deployed an ML framework for real-time IP/UserAgent fingerprinting, integrating classification models to assess threat levels dynamically.
• Researched and prototyped deep learning approaches (CNN) to detect automated connection patterns across web sessions.
• Developed ML-based traffic forecasting models using time series analysis to automate capacity planning and ensure reliable headroom management.
• Designed and deployed real-time monitoring pipelines leveraging statistical modeling and anomaly thresholds to detect performance deviations in traffic flow and system metrics.