Hi there! ▮

Experience

Autonomous Driving Research Intern
NIO NIO（蔚来） (2023.03 - 2024.03)

  • Developed novel “Point-Flow-Point” cyclic architecture for 3D point cloud scene flow estimation
  • Designed base neural network combining GCN, Point Transformer and cross-attention layers
  • Implemented point feature extraction and similarity matching for sparse point cloud scenarios
  • Achieved centimeter-level velocity prediction accuracy in real-road conditions
  • Produced research paper planned for submission to TOMM journal
  • Filed patent application for the proposed technical solution
  • Project successfully concluded with all deliverables completed

Tencent Content Moderation Intern
Tencent Tencent（腾讯） (2024.03 - 2024.12)

  • Developed multimodal content moderation system deployed in production for WeChat Channels
  • Implemented three specialized multimodal models (MMBT-SH, Onepeace-SH, Internvl2.5-SH) for different scenarios
  • Achieved 98% accuracy for safe content identification and 95% coverage for high/medium risk content
  • System processes over 150,000 user uploads daily with 95% interception rate for high-risk samples
  • Lightweight MMBT-SH model (200M parameters) enables efficient moderation with limited resources
  • Onepeace-SH model enhances cross-modal feature alignment for complex content analysis
  • Internvl2.5-SH model combines generative pretraining for detecting implicit violations
  • Technology significantly reduced manual review costs while improving moderation efficiency
  • Currently deployed across Tencent’s core platforms including WeChat Channels

Multimodal Content Moderation Project
Xinhua Xinhua News Agency （新华社） (2024.06 - 2025.4)

  • Developed multimodal intelligent moderation system for Xinhua’s content review platform
  • Implemented specialized models for flag recognition, sensitive symbol detection, and map boundary analysis
  • Achieved 97% accuracy for flag/symbol recognition and 82% accuracy for problematic map detection
  • System covers 5 types of terrorist symbols (ISIS, Nazi, pirate flags etc.) and 62 national flags
  • Designed scene recognition for 4 sensitive scenarios (flag burning, gun violence, etc.)
  • Developed boundary analysis for China’s 34 provincial-level regions with 80% completeness threshold
  • Technology surpassed commercial solutions and large models in accuracy benchmarks
  • Contributed to national network media content security with intelligent moderation solutions
  • System currently deployed in Xinhua’s production environment processing thousands of media daily