• Designed and built an automated infrastructure in Python and C++ for Adobe R&D team to evaluate interactive object segmentation models. Based on that, optimized the process of optimal parameters search for Graph Cut algorithm used by Photoshop.• Designed an innovative Initial Mask based Deep Interactive Object Segmentation algorithm and implemented it on Adobe Sensei platform with OpenCV, PyTorch, and S3 to improve accuracy of Click-based Object Selection Tool for Photoshop and Lightroom 2021.• Designed and implemented a segmentation optimizer module for Photoshop based on the proposed model, fixing errors made by existing segmentation pipeline and generating more fine-grained segmentation results for photos.

• Initiated and participated in the development of an automatic object labeling platform (PaaS) that utilizes user-defined ML models and data from vehicle sensors.• Facilitated the development of CI/CD pipeline with Jenkins + Kubernetes + Gitlab, cutting administration work by 60% and boosting the deployment process by up to 2 - 3 times.• Built and maintained Docker images for Backend-Service and Labeling-Service of automatic labeling platform, increasing the productivity of developers by 1.2 - 1.5 times and reducing the cost of servers by 20%-30%.• Utilized SQLAlchemy + PostgreSQL + Falcon + Redis in Python for the backend of the platform to achieve features including datasets CRUD operations, task monitoring, and metadata management with RESTFUL APIs.• Optimized TF-Serving performance by migrating RESTFUL APIs to gRPC, decreasing labeling time by 70% - 80% and increasing platform productivity by up to 3 - 5 times.• Utilized Petastorm and SparkSQL to filter and save labeled data from Redis to HDFS for following distributed training for different ML models.