Chunghwa Telecom, Software Engineer   Aug 2018 – Present

• Lead the development of iBobby IOS App which can bind with the smart speaker and connect Restful API to allow user to interact with speaker. Achieve over 14000 downloads on the App Store 

• Develop a data warehouse system using Nodejs and MongoDB which stores user’s usage log of iBobby. Analyze user's usage rate in functions to determine which function need to be improved or promoted 

• Develop AI voice filter system using Nodejs’s Express framework, JQuery and Mysql which allows podcasters to acquire their subtitle automatically by uploading the audio file to the system

• Use Pandas to build user data automation system which can efficiently organize data into statistical reports, reducing the time to arrange data by an estimated 1.5 hours/week 

Side Project, Real-Time Facial Expression Recognition and Style Transfer System  Sep 2020 – Dec 2020

• Used image segmentation to detect the mask of user and replace the background with specific picture in real-time 

• Detected facial expression of user with CV2 and mini_XCEPTION. Transferred the style of picture by using MSG-Net based on facial expression of user in real-time

ESun Artificial Intelligence Open Challenge 2020 – NLP Competition 4th Place

• Replaced LSTM with the state-of-the-art technology Bert as the backend to develop classification and name tagging model in PyTorch to extract criminal’s name from the news article

 • Used web crawler to collect data from the web link provided by the organizer and collect other relevant articles as training dataset 

• Deployed model with flask web framework on the Azure platform

Shopee Code League 2021 – Top 30 in Open Category 

• Developed address extraction model, using Bert as backbone to build text sequence labelling model to quickly extract key address elements from unstructured addresses provided by customers

 • Used Pandas and DFS to construct user data classification system which can efficiently identify unique customer from over 500,000 data within 3 hours

ESun Artificial Intelligence Open Challenge 2021 – Image Recognition Competition Top 10%

• Used EffcientNet as backend in Keras to recognize handwriting documents. Built metric learning vectors with Arcface loss to detect unseen words. Classify over 800 words and unseen word with accuracy 90%

 • Through TPU to accelerate the training speed with over five times allows us to try different solutions within the limited game time • Deployed model with flask web framework on the Google cloud platform