[Finetune]Finetuning deep learning model using LoRA technique via Pytorch and PEFT to build a boilerplate filtering model, increase performance by 10x.[LLM Clustering]Design a LLM Clustering Procedure.

[Large Language Model]Make use of LLM to develop an AI Summary feature on Quid Discover Product. - Extensive use of modern LLM tools like Langchain, Langsmith, OpenAI api, and Promptfoo. - Design a sampling algorithm to do representative documents sampling. - Construct the backend API and playground.[Time Series Forecasting]Implement forecasting feature to Trend Detection module. - Survey SOTA paper and find out the best combination of models by experiment. - Pretrain DL forecasting model on our data to boost performance.[Image Captioning]Develop an Image Network POC on Quid Discover Product and Won the first place in our company competition. - Create an end-to-end workflow in a week. Utilize state-of-the-art image captioning method.[Experiment Design]Apply Statistical Methods like ANOVA and Pairwise comparison to verify a better Clustering Algorithm and find out significant best settings all in once and make production code change.[Module Algorithm/Engineering Improvement]Improve Trend Detection and Created a website page for users to play with. The new algorithm is used by Coke, Hershey, and Novartis. - Change to more appropriate statistical tests and introduce new technique in trend scoring. - Apply multi-process and code-refactoring.[Algorithm/ Code Structure Design]Improve Keyword Display Quality for all Dataset - Display Hashtag in names/keywords without interfering the core process. Invent ngram represent notion. - Use lemma as candidates to avoid problems of same word appearing in panel with different forms. Select most frequent token as each ngram's representation. - Avoid unigram contained in multi-gram being displayed solely. Design count filtering logic to achieve the goal.

[Time Series Analysis] [Statistical Analysis]- Invent a Trend Detection methodology from scratch. Currently used by Coca Cola and Novartis. This innovative method has brought in over 100k+ USD profit.[Machine Learning] [Natural Language Processing]The fixed version has been praised by PMs and Sales. They did not need to revise names manually anymore.- Find out systematic flaws of our Cluster Naming algorithm. - Change selection condition to enable unigram as candidates, and apply multigram inclusion technique to make names more meaningful. - Apply part of speech filter. - Design specific input for different dataset based on their characteristic.[Algorithm Design]- Make Ethnicity Distribution Estimation account for minorities to a large extent. The accuracy rate increases over 50% on minorities topics and can better reflect actual Social Media User Distribution.- Improve Algorithm of Generating Representative Words of a cluster by finding a balance between recall and precision.[Experiment]- Design a thorough 3-stage plan to compare performance between tokenizers. Two of the three stages' methods are completely innovative.[Research]- Research on entity embedding methods for related word engine[Guidance]- Prepare an Onboarding PPT for new Data Scientist including Company Tech Stack, Commonly Used Tools, and Benefit. It is then used by other colleagues to introduce our company to new comers.

[AI Application Design]• Developed and Refined the Automatic Text Summarization System for commercial use of a media company • Designed the Interactive User Interface using Flask, JavaScript, Html, and CSS, stored data using SQL• Created sentence compression algorithms with different compression rates

Wroked at Intelligent Agent System Lab, held by Professor Wen Lian,Hsu, former Director of Institute of Information Science, Academia Sinica.[Text Summarization]• Designed an Automatic Text Summarization system beating 10+ popular systems, winning BEST Thesis Award from Taiwan Association for Web Intelligence Consortium• Wrote an Event Extraction package for public use[Paper Writing and Publish]• Published an Automatic Text Categorization paper

[Text Classification]• Refined Automatic Text Categorization method reaching 95% accuracy on 50ks Chinese news data• Conducted experiments on tokenization, word2vec, doc2vec, and fake news detection

[AI Knowledge Translation]• Wrote public Artificial Intelligence articles [AI Research]• Surveyed academical papers related to Artificial Intelligence

[Natural Language Processing]• Improved topic classification model's accuracy by 20%- 30% • Provided proposals for improving topic classification and performing statistical evaluation • Conducted research on topics about machine learning and data science

[Statistical Analysis]• Evaluated different ranking-products methods from statistical perspective[Data Visualization]• Extracted valuable insights from digital reports and made a marketing report[Research]• Shared papers related to statistics on monthly meeting

[Leadership]Earned the Distinguished Club AwardFostered a welcoming atmosphere[Organization Management]Coordinated routine and demonstration meetingsOrchestrated and managed in-club competitionsCoordinated special holiday events, including Christmas and Halloween parties[English Speaking]Won several times of best speaker awards