Machine Learning Researcher · London, UK

Building state-of-the-art large language models (LLMs) that actually work.

As a Ph.D. in Artificial Intelligence recently completed, my research focuses on improving training efficiency and performance of various state-of-the-art deep neural networks(DNNs) including Transformer-based models and convolutional neural networks (CNNS), with a particular emphasis on sound classification tasks, recently publishing the papers related to the topics. Experienced and Skilled expert in various DNNs, and cloud computing (Google and MS Azure) and Linux environment using Git, with exceptional coding and analysis skills. Recently published three deep learning related papers at Neural computing, IEEE, and ICMCL & ICWAPR.

Hyosun Choi

Selected Work

Projects and Papers

01

Novel CNNs with Multi-layer Feature Aggregation for Sound Classification

IEEE SMC 2025 · Novel convolutional neural network architectures using multi-layer feature aggregation and pooling permutations for audio sound classification tasks.

Paper CNN Sound Classification IEEE SMC 2025
02

Dual Representations: Self-Supervised Audio Spectrogram Transformer

Neurocomputing 2025 · A novel variant of Self-Supervised Audio Spectrogram Transformer with dual representations, multi-layer feature fusion, and pooling combinations for sound classification.

Paper Transformer Self-Supervised Neurocomputing
03

The Lottery Ticket Hypothesis and Reinitialisation with Masks

ICMLC 2024 · Investigates the Lottery Ticket Hypothesis applied to deep neural network pruning, exploring reinitialisation strategies with masks to improve training efficiency.

Paper Pruning Lottery Ticket ICMLC 2024
04

Llama 3.3 Domain Fine-tune

QLoRA fine-tuning of Llama 3.3 70B on a domain-specific dataset using Axolotl. Trained on GCP A100 with DeepSpeed ZeRO-3.

PyTorch QLoRA Axolotl GCP
05

Retrieval-Augmented Pipeline

End-to-end RAG system with custom chunking strategies, hybrid retrieval, and reranking. Benchmarked against baseline GPT-4.

LangChain FAISS Python
06

Training Data Pipeline

Scalable data preprocessing and deduplication pipeline for LLM pre-training. Handles 100B+ tokens with quality filtering.

Apache Spark MinHash GCS

Background

Experience

November 2025 — Now

Machine Learning Researcher

Big tech companies · San Francisco, USA

LLM expert, participating in LLM projects from Silicon Valley companies, USA (working remotely)

October 2021 — June 2025

Teaching Assistant

Royal Holloway, University of London · London

Teaching Assistant I've taught major Artificial Intelligence master's subjects,
such as Deep Learning, Natural Language Processing, Autonomous AI systems,
Python Language and so on.
Skills: Associate fellowship of the HEA

2022

Data Science Advisor

gov.uk/Natural England · London

- Worked as a Data Science Advisor for the Artificial Intelligence projects related to generative AI technology
- Worked through Azure GPU cloud computing and development
- Taught how generative AI works to employees more than weekly in Natural England
Skills: PyTorch, Generative Adversarial Networks (GANs), TensorFlow, Artificial Intelligence (AI), Convolutional Neural Networks (CNN), Computer Vision, Microsoft Azure Machine Learning

November 2010 — November 2012

Senior Officer

Standard Chartered Bank Korea · Seoul, Korea

FCU(Finance Control Unit), Finance Accounting Team
- Acted as the Korea representative for conference calls, facilitating communication between the Korea branch and the Standard Chartered Group.
- Led the improvement project for the newly adopted K-EDW system(domestic financial accounting system) : testing, maintaining, or launching software products, and experience with software design and architecture.
- Managed month-end and year-end closing processes for financial statements at Standard Chartered Bank Korea.
Skills: Java · Data Science · Programming · Software Project Management · Software Development · Software Design.

February 2005 — September 2007

Engineer

Samsung SDS · Seoul, Korea

Software Engineer, Sec Global improvement part , Sec Global IS Team, Samsung Electronics division
- Managed the Worldwide Trading Network (WTN) System for Samsung Electronics, supporting the Global Finance Team through ERP: testing, maintaining, or launching software products, and experience with software design and architecture.
- Implemented and maintained the SAP ERP system (FI, SD) using ABAP language for overseas subsidiaries, providing training on accounting manuals.
- Successfully installed the WTN system on the Samsung Austria server and presented the system manual during a business trip to Austria.
Skills: ABAP, SAP ERP, SAP FI consultant, Sun Certified Java Programmer, Financial Analysis, Accounting, Data Analysis

Background

EDUCATION

March 2020 — December 2025

Ph.D. Artificial Intelligence in Computer Science

Royal Holloway, University of London · London

Improving training efficiency and performance of various state-of-the-art deep neural networks(DNNs) including Transformer-based models and convolutional neural networks (CNNS), with a particular emphasis on sound classification tasks

September 2018 — December 2019

MSc. Machine Learning in Computer Science

Royal Holloway, University of London · London

Top level of GPA
Master's Thesis: The Experiments of DCGANs and WGANs in Generative Deep Learning

February 2008 — August 2010

MBA, Master of Business Administration

KAIST (Korea Advanced Institute of Science and technology) / Techno MBA · Seoul, Korea

GPA 3.73/4.3 (93.66/100)
Master's Thesis: ‘Storytelling Marketing' Analysis and Strategy

March 2000 — February 2005

Bachelor of Science, Computer Science and Engineering

Ewha Womans University · Seoul, Korea

GPA 3.62/4.3 (92/100)
Honor student of Ewha Womans University
English study in Boston, U.S.A.: Boston Univ. CELOP(Language School) in the U.S.A.

Technical

Skills

ML & Training

  • PyTorch / JAX
  • HuggingFace Transformers
  • LoRA / QLoRA / PEFT
  • DeepSpeed / FSDP
  • RLHF / DPO

Infrastructure

  • GCP / AWS
  • Kubernetes
  • Docker
  • Terraform
  • Weights & Biases

Languages

  • Python
  • CUDA / Triton
  • Bash
  • SQL

Let's talk.

Open to interesting problems in LLM training, alignment, and production ML. Feel free to reach out.