Shicheng Chen Email & Phone Number

Singapore

Singapore

Post Recommender System and Search Engine-Built the company's first recommender system for fashion images, achieving a 20% performance improvement over AWS Personalize.-Developed the company's first search engine using large multimodal models, surpassing the keyword-based baseline by 44.9%.-Led the collaboration of multiple teams to architect a distributed system using serverless functions for lightweight application logic, coupled with blob storage, SQL/NoSQL databases, and vector… Show more Post Recommender System and Search Engine-Built the company's first recommender system for fashion images, achieving a 20% performance improvement over AWS Personalize.-Developed the company's first search engine using large multimodal models, surpassing the keyword-based baseline by 44.9%.-Led the collaboration of multiple teams to architect a distributed system using serverless functions for lightweight application logic, coupled with blob storage, SQL/NoSQL databases, and vector stores for diverse data operations, leveraging managed ML infrastructure for compute-intensive workloads, achieving sub-second response times at scale.Bra Recommender System-Optimized mobile pose estimation model to achieve real-time performance (30 FPS), reducing mean endpoint error by 70% compared to Google's MoveNet.-Trained a semantic segmentation model that extracts specific human parts and reduces segmentation errors to less than 1% in complex background environments.-Built neural networks for 3D human body reconstruction using a synthetic dataset, cutting processing time by 80% compared to previously released algorithms while maintaining accuracy.-Integrated these models into Victoria's Secret and Verifyt apps, used by over 100,000 users.-Architected an MLOps pipeline for continuous model improvement, deploying a Flask web server in Kubernetes (K8S) for interactive data collection, labeling, and model retraining. Show less

Singapore

Hand Pose Registration:Proposed a kinematic chain-based hand pose registration layer.Achieved a 27% reduction in mean end-point error compared to prior work on synthetic image enhancement.Published findings in the International Conference on Computer Vision (ICCV) 2021.Semi-Supervised Deep Learning for 3D Hand Pose Annotation:Developed a semi-supervised deep learning model for annotating 3D hand poses and meshes.Labeled an 80,000-image dataset captured by 4 RealSense… Show more Hand Pose Registration:Proposed a kinematic chain-based hand pose registration layer.Achieved a 27% reduction in mean end-point error compared to prior work on synthetic image enhancement.Published findings in the International Conference on Computer Vision (ICCV) 2021.Semi-Supervised Deep Learning for 3D Hand Pose Annotation:Developed a semi-supervised deep learning model for annotating 3D hand poses and meshes.Labeled an 80,000-image dataset captured by 4 RealSense depth cameras, achieving a within 4mm mean end-point error with only 1.5% manual annotation.Published findings in the British Machine Vision Conference (BMVC) 2021.Apache Singa & ONNX Integration:Collaborated with a team of 5 engineers to extend Apache Singa's capabilities to support the ONNX model standard.Enhanced Singa's automatic differentiation engine to accommodate additional neural network layers.Successfully merged pull requests into the main branch.GPU Memory Optimization in Apache Singa:Introduced and optimized a GPU memory management module.Achieved a 25% reduction in memory usage with a minimal increase of approximately 3% in computation time.Applied optimizations for popular neural networks, including ResNet, UNet, and VGG. Show less

Singapore

Graph Embedding for Road Graphs:Engineered a neural network layer specialized in graph embedding to mine neighborhood information of road graphs.Achieved superior performance over 6 established methods, including DL-STF and FMSTA.Published research findings in the SIAM International Conference on Data Mining (SDM) 2019.Spatial-Temporal Convolutional Neural Network for Taxi Dispatch:Designed a convolutional neural network incorporating spatial and temporal awareness for… Show more Graph Embedding for Road Graphs:Engineered a neural network layer specialized in graph embedding to mine neighborhood information of road graphs.Achieved superior performance over 6 established methods, including DL-STF and FMSTA.Published research findings in the SIAM International Conference on Data Mining (SDM) 2019.Spatial-Temporal Convolutional Neural Network for Taxi Dispatch:Designed a convolutional neural network incorporating spatial and temporal awareness for an optimized taxi dispatch system.Outperformed 5 baseline methods, notably ST-ResNet, particularly in high-demand scenarios.Published research findings in the Database Systems for Advanced Applications (DASFAA) 2019. Show less

Master'S Degree, Computer Science

Bachelor'S Degree, Computer Software Engineering