Taeyang Kim
Machine Learning Engineer, Pattern Inc.
Actions
Taeyang Kim is a dynamic Machine Learning Engineer and Computer Scientist with a passion for innovation and a track record of excellence in the field of data science and artificial intelligence. He studies Master of Science in Computer Science at the Georgia Institute of Technology, building upon his strong foundation from Brigham Young University (BYU), where he will graduate with a Bachelor of Science in Computer Science, emphasizing Data Science.
During his time at BYU, Taeyang distinguished himself both academically and through leadership roles. He was the recipient of a Full Academic Scholarship for the 2021-2022 academic year and served as the Tech President for the BYU Korean Business Student Association. His commitment to excellence is further demonstrated by his remarkable achievements in various competitions:
1st Place in the 2025 Utah AI Hackathon Competition (1st out of 186)
1st Place in the 2024 Utah Developer Hackathon Competition - Justbuild.ing
1st Place at the 2024 BYU Computer Science Senior Capstone Project Competition in Machine Learning
1st Place at the 2024 BYU ACM YHack Hackathon Competitions, earning the prestigious Benny Award
1st Place at the 2023 BYU ITCSA Raspberry Pi Competition, recognized as the best robotics competition at BYU
1st Place at the 2022 BYU ACM YHack Hackathon Competitions
1st Place at the 2017 DIPA Internet of Things Hackathon Competition in South Korea
Professionally, Taeyang has made significant contributions in the tech industry:
At Pattern (Apr 2023 – Present), he serves as a Machine Learning Engineer in Lehi, UT. Notably, he was honored as Pattern’s 2024 "Employee of the Year" out of 1,700 employees. His innovative work in automating IT ticket resolutions using advanced AI models drastically reduced ticket numbers by 75% and resolution time by 87%.
With Converus (Jun 2023 – Sep 2023), he developed sophisticated web applications utilizing GraphQL, Firebase, and modern JavaScript frameworks.
At Podflow (Mar 2023 – Apr 2023), he spearheaded the deployment of real-time audio editing software, enhancing text-to-speech and speech-to-text conversion capabilities.
As a Research Assistant at the BYU DRAGN Lab (May 2022 – Apr 2023), Taeyang developed a full-stack web application for an Amazon Alexa Competition finalist LLM model and invented a multifunctional farming robot integrating advanced sensors and robotics.
His earlier roles include software engineering positions at Leftovers and BYU Broadcasting, where he showcased his skills in full-stack development, database management, and optimizing CI/CD pipelines.
With a blend of academic excellence, professional expertise, and innovative project experience, Taeyang Kim brings a wealth of knowledge and a fresh perspective to the field of machine learning and data science. His contributions have not only advanced technologies within his organizations but have also set the stage for future innovations in the industry.
Area of Expertise
From Model to Market: Deploying Generative AI Video & Audio Systems at Scale
How do you take a generative AI idea and turn it into a cloud-native, scalable product? In this talk, I’ll walk through the full-stack journey of deploying a SaaS platform for generating marketing content—focusing on audio and video use cases.
As project lead, I oversaw a cross-functional team building an asynchronous, event-driven architecture capable of handling thousands of concurrent generation tasks. The system integrates FastAPI, React, and modern cloud services, leveraging Terraform, Docker, and AWS Fargate.
You’ll learn how we deployed custom generative models using tools like Replicate and WAN 2.1, and designed a modular backend pipeline to orchestrate the workflows—alongside real-time feedback for users.
This session will be especially useful for ML engineers and MLOps professionals looking to bridge the gap between model development and product deployment.
LLM Network Intrusion Detection System
Current network intrusion detection systems (NIDS) struggle to keep pace with the sophistication and evolving nature of cyberattacks. Traditional signature-based and rule-based systems are often brittle and easily bypassed, while anomaly-based systems suffer from high false positives. This leaves an alarming gap in network security, exposing organizations to data breaches, financial losses, and reputational damage.
We propose the development of a full-stack Large Language Model (LLM) product designed to revolutionize threat analysis through the comprehensive examination of system and network logs as natural language. This solution aims to provide systems that can understand complex attack protocols, detect novel attacks, and adapt to evolving threats. Our system enhances security with proactive and comprehensive defense against a wider range of threats, improves efficiency through accurate threat detection and reduced false positives, and offers a future-proofed defense with continuous adaptation to evolving threats, ensuring long-term effectiveness and protection against emerging attack vectors
Knowledge Graph–Driven Feedback Loops to Optimize TikTok Shop Video Marketing
###Background
TikTok Shop has rapidly become one of the most important platforms for e-commerce growth, especially among Gen Z consumers. According to recent market data, over 40% of Gen Z shoppers prefer discovering and purchasing products directly on TikTok, surpassing traditional search engines or brand websites. As a result, e-commerce brands are shifting significant attention and ad spend toward TikTok Shop, where short-form video content is directly tied to conversion outcomes.
Despite this, most social-media marketers still rely on vanity metrics—likes, shares, and follower counts—which often fail to reflect true sales impact. With TikTok Shop now offering video-level sales and GMV data, brands can for the first time create a closed-loop marketing system where creative choices can be evaluated and optimized based on actual revenue performance.
###Objective
To create and evaluate a knowledge-graph–based framework that links TikTok Shop video attributes (hooks, camera shots, script phrasing, voiceovers), audience engagement, and gross merchandise value (GMV). We hypothesize that providing marketers with these model-driven insights will enable data-informed content optimization and yield a significant increase in per-video GMV.
###Methods
We will leverage TikTok Shop seller analytics to extract video-level data—hook types, camera shot categories, script phrasing patterns, and voiceover styles—alongside engagement metrics (views, watch time) and sales outcomes (GMV, units sold). These elements will be represented in a Neo4j-based knowledge graph, linking videos, products, and audience segments.
To conduct a deep dive into each video’s creative and performance elements, we will use advanced LLMs and multimodal Vision models to analyze video content frame-by-frame. These models will automatically identify and categorize visual, audio, and narrative features, and convert them into structured GraphDB elements, enabling deeper semantic analysis and insight generation.
We will test this framework at scale using real TikTok Shop brand data. Pattern Inc. manages over 200 e-commerce brands, with more than 20 actively selling through TikTok Shop. These include well-known names such as Philips, Optimum, and Thorne—all with verified GMV and TikTok sales data. Additionally, we will test our system with a smaller cohort of brands to validate performance across brands at different maturity levels.
We will evaluate pre- vs. post-system performance (including data extraction, modeling, insight delivery, and dashboard interaction) using paired t-tests and time-adjusted regression models.
###Results
Results are still pending. We expect the execution of this study to be complete by the end of June 2025. We anticipate that our LLM-powered, knowledge graph–driven feedback system will significantly improve the ability of brands to interpret TikTok video performance, adjust creative strategies, and increase GMV per video.
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top