Our team is working on QuickList: https://www.cnbc.com/2023/09/13/amazon-rolls-out-generative-ai-tool-to-help-sellers-write-product-listings.htmlDeveloped a full-stack internal tool that connects to DynamoDB and Bedrock, Amazon's latest LLM model.Use React and Typescript as frontend, Java as backend, RESTful API, and DynamoDB as the database.

Develop CLEAN frontend using Typescript and Angular. https://clean.frontend.mmli1.ncsa.illinois.edu/configurationDeveloped PlasmidMaker, a full stack application, for iBioFab using Django, sqlite3, and React.js.

Developed recommendation algorithm for retail products by utilizing light GBM, and XGBoost.Optimized recommendation model to boost inference time by 32% by utilizing parallel programming on GPUDeveloped a prototype to automate the workflow of measuring store area by utilizing SLAM and Microsoft KinectDeveloped an automated product tagging model prototype to boost efficiency and reduce human power by utilizing normalized google distance, cosine similarity, and CKIP transformers (ALBERT, BERT, and GPT2)

Designed deep learning structures to identify dog breeds and individual dogs and achieved over 85% precision and recall on both Stanford Dogs dataset and real-world user dataset with over 5k imagesVisualized the model learning results by applying gradCam and SHAP on testing data (image)Implemented style transfer deep learning models with AnimeGanv2 using real-world user dataWorked with Prof. Hung-yi, Lee (NTU), reported audio recognition progress, and discussed solutions in monthly meetingsDesigned models using both transformer-based and convolution-based networks to solve cross-device audio recognition tasks on real-world data and achieve over 90% precision and recallUtilized self-supervised learning and transformer models like SSAST and DistilHubert to achieve over 80% precision and recall on cross-devices audio recognition tasks on real-world user data with over 40k sound files