Experience

• 
  PDD Holdings

  Shanghai, China

  A global e-commerce platform (Nasdaq: PDD), parent company of Pinduoduo and Temu.

  Software Engineer, Machine Learning Infrastructure

  Jul. 2023 - Nov. 2025

  I worked on offline training infrastructure for Pinduoduo's e-commerce search and recommendation systems, serving 800M+ users. My work focused on distributed training frameworks, sparse embedding systems, and end-to-end training performance optimization. I also collaborated closely with algorithm teams to improve training workflows and diagnose model and training-system issues in production.

  Scalable Training for Generative Recommendation Models

  • Implemented training workflows for generative recommendation models (HSTU) using PyTorch and TorchRec, integrating DDP and DMP to scale billion-parameter sparse training on 300M+ DAU datasets for multi-task workflows (recall, CTR, CVR).
  • Built a pipelined training execution framework that overlaps CPU-side data preprocessing, sparse embedding lookup, forward/backward computation, and parameter updates, increasing training throughput by 40%+.
  • Optimized sparse embedding communication by deduplicating feature IDs before TorchRec embedding-table lookups, reducing all-to-all traffic and improving end-to-end throughput on 32 GPUs by 25%+.

  Training Framework Migration

  • Migrated a recommendation model training pipeline from TensorFlow to PyTorch on multi-node H100 clusters.
  • Integrated a custom C++ sparse Parameter Server as a CPU backend to offload massive read-only embedding tables, effectively bypassing GPU memory limits for memory-intensive downstream models.
  • Rebuild the end-to-end training lifecycle including sharded checkpointing, high-throughput data prefetching, and cross-framework consistency checks to match existing training workflows.
  • Optimized training efficiency through data prefetching and CPU–GPU coordination, leading to 30%+ higher GPU utilization and 15% faster training compared to the legacy TensorFlow system.

  Semi-Synchronous Training Mechanism

  • Extended SyncReplicaOptimizer (TensorFlow v1.0) with a semi-synchronous mechanism, enabling non-blocking worker execution to improve system throughput while preserving convergence stability.
  • Implemented token-based gradient admission and staleness control on the parameter server to aggregate gradients at a fixed global batch size while filtering stale updates.
  • Improved training throughput by 60%+ over native synchronous training on a recommendation CVR model, reaching 85% of asynchronous efficiency with a +0.3% AUC improvement.
  • Supported parameter-level update strategies via TensorFlow graph modifications to meet model-specific requirements.

  Distributed Training Coordination System

  • Designed a coordination mechanism for multi-node startup synchronization and runtime metadata distribution.
  • Implemented leader election (ZooKeeper + CAS) to assign a primary worker which handles code distribution and cluster metadata collection, eliminating external dependencies during job startup.
  • Coordinated worker readiness and generated distributed configurations to enable reliable cluster bring-up.

  Machine Learning Platform Backend

  • Contributed to backend development of an ML platform, supporting task submission, scheduling, and monitoring.
  • Developed RESTful APIs (Python/Java) for training jobs, resource management, and user access control.
  • Unified inconsistent API endpoints across services by implementing a proxy-based routing layer, and validated consistency by comparing proxy and direct API responses in production.
• 
  SenseTime

  Shanghai, China

  An AI company (HKEX: 0020) specializing in computer vision and deep learning.

  Software Engineer Intern

  Jul. 2022 - Sep. 2022

  I worked on backend and infrastructure for an enterprise deep learning platform, with a focus on backend services, Kubernetes-based deployment, and model training workflows.

  • Developed Golang (Gin) backend services and REST APIs for an enterprise deep learning platform, supporting algorithm management, training workflows, and access control.
  • Built and maintained Helm-based deployment workflows for Kubernetes services and training jobs, including chart validation, environment configuration, and release deployment.
  • Integrated pretrained models and training pipelines into the platform, implementing model configuration validation and supporting model artifact management with object storage.
• 
  Pivot Studio

  Wuhan, Hubei, China

  A student technology club building digital products and services for the university community.

  Software Developer

  Oct. 2021 - Apr. 2022

  Developed and maintained backend services for 1037 Tree Hollow, an anonymous campus social platform serving thousands of students at Huazhong University of Science and Technology.

  • Built backend features and REST APIs using Go and MySQL.
  • Supported ongoing service maintenance and refactoring to keep the platform reliable and maintainable.