Projects 🏭

2022 E-Commerce Visual Language Pre-training

• Developed a large-scale (100M pairs) e-commerce visual-language pre-training model for enhanced product understanding.
• Proposed BagFormer, a dual encoder for cross-modal retrieval using bag-wise interaction and entity-aware text granularity. Achieved top recall with low latency.
• Ranked 1st in MUGE Challenge with BagFormer.

2021 WeChat Video Product Identification

• Led design & development of scalable video product ID pipeline: data processing, model training, inference acceleration, and deployment (TensorRT, Kubernetes, Pulsar).
• Led the development of major modules: multimodal video classification(ResNet-50, BERT, and NextVlad), product detection(YOLO-v5), product retrieval(Siamese Network), and product tagging(ALBEF+XGBoost).
• Boosted overall GSB by 12.37%, Top-3 GSB by 26%, contributing to WeChat Video reaching 500M DAU.
• Won 2nd place in the AAAI 2021 Watch and Buy Challenge.

2020 WeChat Pay Merchant Understanding

• Tagged 7.7M merchants with 370M high-quality tags using distant supervision and active learning.
• Modeled merchant-tag relations as a reading comprehension task using BERT + fine-grained entity types.
• Achieved 96.6% top-1 accuracy on head merchants and 89.0% overall.
• Improved pickup rate by 6.58% and usage rate by 0.28% (via A/B testing).

2019 WeChat Search Query Improvement

• Built relevant query recommendation system to improve search satisfaction.
• Constructed co-click query graphs (Hadoop), filtered with FastText, trained DSSM with hard negatives, and deployed FAISS PQ Index for real-time similarity search (latency < 100ms).
• Led to 4.5% daily query view increase, raising search ad revenue by 3%.