A Data Ingestion Pipeline that Preprocesses Raw Events and Publishes Them to Apache Kafka in Real-time• Saved 50% of machine resources and 20% of redundant traffic in one upstream consolidation• Tuned the JVM-based pipeline to scale with the volume of user instrumentation events at Yahoo• Enriched data with Redis caching for downstream use cases, i.e., feature computation, user profiles, and analyticsHulkbuster - A Recommendation Reasoning Pipeline that Badges Articles for Users• Delivered badges such as Most Read, Most Discussed, You Follow, and New Topic based on ML features.• Served as the Engineering Lead in collaboration with a group of Product Managers, Designers, and Editors.

Leopardon - A Model Serving Automation Pipeline for Real-time Content Recommendation System• Raised model velocity to 100% automation by helping Data Scientists launch online experiments in one go• Enabled continuous deployment of retrained models for online A/B testingFine-grained Feature Design and Feeding for Real-time Precision Ranking• Introduced collaborative filtering features for precision ranking on top of recalls in a multi-phase ranking system• Alleviated user cold start problem with more convoluted user/document features

New Feature Design and Implementation on Yahoo Content Recommendation and Personalization Engine• Designed an article-level score to exploit redirected page views as trending signals outside of Yahoo network• Established the full cycle of a real-time pipeline from data logging, feature computation, to re-training model• Maintained 94% or above code coverage with JUnit in every codebase

User Intention Understanding on Taiwan Open Platform for Educational Resources– Revealed search keyword trends on the website with millions of entries of user log data– Developed a keyword generation pipeline for personalized resource recommendation in primary educationSQUAT - a Sequencing QUality Assessment Tool in Bioinformatics– Tools released on Github and academic work admitted to BMC Genomics, a peer-reviewed scientific journal– Provided read mapping analytics with visualization in a portable HTML report– Devised metrics for researchers to identify poor-quality data upon terabytes of data gatheringhttps://github.com/luke831215/SQUATDe Novo Genome Assembly– Proposed an ML framework of genome assembly with XGBoost for scoring quality reads– Achieved the relative improvements of assembly results from 25% to 35% on N50 statistic– Enhanced genome contiguity with less input overhead and sequencing read required

Compulsory Military Service

Pattern Recognition– Simulated patterns of malicious mail attacks with log-based data hashing and clustering– Reduced volume by designing a white-listing mechanism to filter out known legitimate entries