• Developed, tested, deployed, and monitored visually appealing features for the SQL worksheets in Snowflake’s new UI, Snowsight, using TypeScript; fixed UI bugs and collaborated closely with UX designers to ensure an optimal user experience.• Added Datadog telemetry to both frontend and backend to analyze feature usage and set up automated crash alerts, working with data scientists to enhance monitoring and insights.• Improved the Go backend API service by refactoring endpoint calls to a more REST API format to redirect traffic. Enhanced performance monitoring and implemented advanced crash monitoring for better incident investigation and resolution.

• Engineered a Streamlit-based web application providing real-time analysis and visualization of user-input data, leveraging machine learning techniques for insightful data interpretation.• Employed the Snowpark Python API for managing data operations, such as creating tables and executing queries.• Collaborated with the design team on polishing the UX, implementing constructive feedback from internal users to optimize the user experience.

Open-Source Project Mirror3D • Designed a human-in-the-loop workflow that enables efficient iterative annotation of mirror 3D planes.• Engineered an intelligent mirror image annotation tool facilitated by a pre-trained ResNet-50 classifier; constructed asemi-automated 3D plane annotation interface using Open3D that allows inspection of a 3D mirror plane estimate.• Created a dataset with 7,011 annotated 3D mirror planes in 5,894 RGBD frames.• Built an efficient web interface based on HTML to help verify 2D object segmentation and 3D plane annotation.3D Scene Reconstruction from Panoramic Images• Programmed a Three.js-based web interface that allows users to edit a 3D scene by applying materials on a model.• Built and maintained a full-stack 3D scene annotation task management application to load and browse 3D scenes.• Implemented features to allow user registration and login for crowdsourcing support.• Presented a literature review on single-view shape reconstruction and identified the strengths and weaknesses.• Developed learning-based approaches for image-based shape retrieval to learn a joint image-shape embedding space using ResNet-FPN and PointNet++.• Developed parallel shape features computing using Python multiprocessing library; optimized data loading process using HDF5 file format and achieved 2x faster IO speed.• Benchmarked the performance based on the chamfer distance, normal consistency, and point-wise reconstruction F-score; demonstrated the advantages of different methods and identified areas for improvement.• Developed 3D model headless rendering using Python Blender; optimized the process using GNU parallel and boosted rendering by 20 times.• Developed two image-based approaches for 3D shape pose estimation using light field descriptor, with 7% higher accuracy than another popular method, PoseFromShape.