Seonglae Cho

Experience

Kakao Mobility (software engineer)  Seoul, South Korea

• Led a 3-person team in developing an app for 3D maps as a part of the Autonomous Driving pointcloud data pipeline.

• Designed an API protocol of automation features in collaboration with the AI team to integrate ML into the service.

• Structured and managed the Google GKE namespace to ensure compatibility with local Kubernetes, facilitating seamless development.

Stryx Seoul, South Korea

• Reduced build time by 70%, simplified dependency management by merging multiple repositories into a mono-repository.

• Achieved OS independence by refactoring a C++ 3D algorithm module to Rust, using Node.js NAPI binding.

Stryx Seoul, South Korea

• Enhanced internal server efficiency by cutting bandwidth and TTFB by 80%, introducing multi-layer Redis caching.

• Elevated team productivity through downsizing the Docker image from 2GB to 180MB, utilizing multi-stage builds in CI.

Academic Paper

RTSUM, Yonsei University Data & Language Intelligence Lab Mar 2023 – Aug 2023

Contest

GenAI Competition at Yonsei University in 2023                                     Seoul, South Korea

• Won the Grand Prize in collaboration as a part of a team of two for the project MBTI-GPT. Dec 2023

Projects

TPCxAI, Yonsei University                                                                   Mar 2024 – May 2024

• Implemented a Big Data system for the AI pipeline from scratch on AWS using Hadoop and Spark cluster, employing HDFS.

• Migrated the manual EC2 cluster to Amazon EMR, reducing management costs by simplifying system dependencies.

ReSRer, Yonsei University Data & Language Intelligence Lab               Sep 2023 – Jan 2024

• Improved ODQA(Open-Domain Question Answering) performance by 20% with zero-shot LLM context manipulation.

• Enabled large-scale QA benchmarks by indexing 21M Wikipedia passages into a Milvus vector database in 12 hours.

• Boosted LLM evaluation by 40% by introducing a multi-GPU local inference server, Huggingface TGI with asynchronous batch processing.

MBTI GPT, AI Service with 1000+ users                                        Oct 2023 – Feb 2024

• Implemented enterprise-level application of AI personality analyzer using Redis, OpenAI API, Node.js and Faiss.

• Ensured the security of RAG by the approach of real-time vector indexing with dynamic data splitting in-memory.

• Reduced OpenAI API costs by 30% by prompt optimization, utilizing code from ‘LLM as optimizers’ paper.

Yokhal                                                                                                 Mar 2024 – Apr 2024

• Accelerated LLM training speed using Pytorch’s multi-node distributed training DDP and FSDP with 4 x RTX3090.

• Implemented a user-centric Chatbot by fine-tuning Gemma on Korean chat and wiki datasets on PEFT QLoRa.

LLaMa2GPTQ                                                                                       Jun 2023 – July 2023

• Designed and developed a local assistant AI application utilizing LangChain, ChromaDB and Streamlit.

• Optimized computing and memory cost by 75% using 4-bit GPTQ quantization applied to the LLaMa2 model.

Texonom                                                                                          Nov 2021 – Jun 2023

• Integrated a Recommender system into the service, devised to build a context extended model through ONNX.

• Developed vector search API for RAG by embedding whole 30,000 pages in service into Postgres pgVector database.

To smooth                                                                                                Jun 221 – July 2021

• Extended Chaikin’s Smooth Algorithm implementation to a multi-dimensional library and deployed it to the NPM registry.

HDGen, Kakao Mobility                                                               Oct 2020 – Sep 2022

• Improved vector mapping speed by developing a method to extract normal vectors from pointcloud data.

• Rendered millions of 2D/3D vector data and pointcloud points within the app manipulating Geoserver and PostGiS.

MMS-TWR, Stryx                                                                                   Jun 2020 – Dec 2020

• Upgraded internal service reliability by migrating from bare metal to OpenStack by building Docker image for applications.

MMS-VHCL, Stryx                                                                                  Jun 2020 – Oct 2020

• Obtained LiDAR sensor control for vehicles with real-time pointcloud communication from hardware.

STPano, Stryx                                                                                        Mar 2020 – Jun 2020

• Raised the precision of drone coordinate positioning by developing a mathematic algorithm for panoramic images.

Education

YONSEI UNIVERSITY                                     Seoul, South Korea

• Expected GPA: 3.68/4.5 (92.1%), Jinri Scholarship (1 semester of 2018)

• Key Modules: Reinforcement Learning, NLP, Big Data, ML, Computer Vision, Computer Security, Signal Processing, Algorithm for AI, Automata, Computer Architecture, OS, Linear Algebra, Statistics, Discrete Math, OOP, Data Structure

UNIVERSITY OF CALIFORNIA, RIVERSIDE                            Riverside, CA, United States

• Relevant courses: Capstone Design, Computer Graphics, Compiler Design

Activities and Leadership

Yonsei University                                                                                       Seoul, South Korea

• Organized events such as retreats and coordinated several cooperative purchases for the department’s shared uniforms.

University of California, Riverside                                              Riverside, CA, United States

• Extracurricular activities: Member of the Tennis Club, visited CES 2023, and observed a Falcon 9 launch event in Florida.

Jaramter Kindergarten                                                                            Masan, South Korea

• Conducted activities with disabled children during each vacation throughout middle and high school.

Skills and Interest

• Academic Interests: Explainable AI, AI Question Answering, Transformers, LLM, Mechanistical Interpretability

• Technical Writings
• Cho, S. (2024, April 10). Superposition Hypothesis for Steering LLM with Sparse AutoEncoder. Medium. https://seongland.medium.com/superposition-hypothesis-for-steering-llm-with-sparse-autoencoder-c07b74d23e96
• Cho, S. (2024, April 22). Reversing Transformer to Understand In-Context Learning with Phase Change & Feature Dimensionality. Medium. https://seongland.medium.com/reversing-transformer-to-understand-in-context-learning-with-phase-change-feature-dimensionality-13cbf8a2f984

• Programming Languages: Python, Rust, Typescript, C++, JavaScript, Bash, Java

• Skills: Pytorch, PEFT, GCP, FSDP, Model Training, Prompt Engineering, RAG, Vector Database, Hadoop, Faiss, AWS PostgreSQL, Redis, Cloud Computing, Kubernetes, Docker, ONNX, Git, CI/CD, Distributed Systems, ETL, Spark

• Languages: English (Fluent), Korean (Native), Japanese (Intermediate)