Bellevue, Washington, United States
CPP AI Training and Label Discovery Team
- Worked in Classification and Policy Platform (CPP) AI Training and Label Discovery Team - Build self-service ML model training system using active learning techniques - Contributed to an LLM-based assistant that enables users to build ML models or keyword rules with minimal ML expertise. - Designed and implemented image upload support for the assistant, enabling users to upload product images and retrieve similar ASINs from the Amazon catalog. - Implemented flexible model checkout functionality for early model finalization based on metrics, improving deployment turnaround. - Designed and built a noise reduction workflow to filter inconsistent or low-quality user labels, improving training data quality.
- Worked in Classification and Policy Platform (CPP) AI Inference Team. - Build scalable ML Infrastructure for over 10000 ML models for Product Classification in Amazon catalog in cost effective and efficient way. - Implemented a custom asynchronous inference platform, onboarding 200+ ML models and achieving annual inference cost savings of over $1M. - Designed and developed the Classification Result Detection System to monitor and identify anomalous classification results using Amazon Kinesis Firehose, AWS Lambda, SNS, and Athena.
- Containerized the machine learning backend on an EC2 instance and deployed it to AWS SageMaker, thereby simplifying model deployment, enhancing resource isolation, and leveraging SageMaker's built-in features for cost-effective and secure model management. - Configured auto-scaling features to dynamically adjust resources, significantly enhancing system scalability and cost-efficiency - Conducted load testing on the SageMaker Endpoint, identifying bottlenecks and implementing optimizations to improve system availability and reliability.
- Worked on class auto-generation and auto-assignment of Presence.fit, an AI-powered fitness app in NodeJS - Developed 10+ GraphQL APIs of video tagging, personal workout structure update and class generation. - Realized the feature that can automatically assign users training classes based on their time availability and goals. - Created the class generation module that can generate a training class by automatically stitching the clip videos focusing on specific target muscles area.
- Developed and maintained SAP Work Zone, a cloud-based digital workplace in Ruby and JavaScript. - Created 30+ new OData APIs for advanced operations of users’ data for Work Zone mobile app and recommendation system. - Fixed 10+ backend and frontend defects of SAP WorkZone including data issues and display issues. - Implemented 5+ feature improvements for groups recovery to make it more convenient for restoring group data. - Utilized text-to-speak tools of Windows and macOS to improve the frontend accessibility.
- Proposed and built a multi-scale non-local network with Keras to achieve automatic EEG signal detection to detect epilepsy. - Conducted experiments to evaluate the performance, achieving accuracy classification performance of more than 94.1% which is higher than other recent networks. - Published paper MNL-Network: A Multi-scale Non-local Network for Epilepsy Detection from EEG Signals.