Summary

Machine-learning researcher and engineer specializing in multimodal pre-training, text-image alignment, post-training, model evaluation, and production systems. Led foundation-model projects and built distributed training and evaluation infrastructure across the model-development lifecycle.

Resume downloads

Open source

OpenLocalAgent: Building small tool-calling agents from scratch in pure PyTorch using open data, with pretraining, midtraining, post-training, and evaluation across ten public agent benchmarks. Developed dataset-level loss analysis to investigate differences in predictability across coding, structured, and web-text corpora.
OCR: Building a reproducible evaluation and improvement framework for sub-1B document VLMs, covering accuracy, calibration, robustness, and LoRA adaptation. Unified public document datasets into a shared evaluation and training format, with deduplication, held-out splits, task-specific metrics, and ablation studies.
Absorbed: Developing an open-source SwiftUI roguelike with turn-based combat, branching expeditions, spirit capture, persistent progression, and English/Korean localization. Built deterministic content, asset, localization, and balance checks; currently a playable iOS prototype under active development.
Hugging Face ecosystem contributor with 40+ contributions, including 28 public pull requests authored in Transformers across model architectures, processors, conversion scripts, distributed training fixes, tests, and documentation.
Hugging Face Transformers contributor: led the addition of Segment Anything 2 (SAM2) support to huggingface/transformers.
- Implemented and refined image/video segmentation model support, processors, documentation, conversion flow, and integration tests through a long-running community review cycle.
- PR: Add Segment Anything 2 (SAM2); merged 2025-08-13.
- Documentation: SAM2 model docs, credited as a model contributor.
Hugging Face Transformers contributor: opened Add Molmo2 and published danelcsb/Molmo2-4B on the Hugging Face Hub.

Education

M.S. in Electrical Engineering and Entrepreneurship, Korea Advanced Institute of Science and Technology, KAIST, 2020.03-2021.02
B.S. in Electrical Engineering, Pohang University of Science and Technology, POSTECH, 2015.03-2020.02
B.S. in Electrical and Computer Engineering, University of Illinois, Urbana and Champaign, UIUC, 2018.01-2018.12

Work experience

2026/01-Present: Data Scientist
- Toss Bank (토스뱅크)
- Building on-premise tool-calling agents that route requests between direct question answering and RAG over internal documents retrieved with image embeddings.
- Building an in-house language model benchmark aligned with public evaluation protocols and scoring heuristics to compare Qwen and GLM models and analyze performance differences.
- Developing mobile ID card recognition and facial verification with ONNX models running through WebGPU; combining face anti-spoofing, multitask depth/normal/keypoint estimation, and face-embedding similarity.
- Post-trained and evaluated an approximately 1B-parameter vision-language model in an end-to-end document extraction pipeline, reaching 61% exact-match accuracy for automation-ready outputs.
2021/09-2026/01: Machine Learning Engineer
- SuperbAI
- Held approximately one year of team-leadership responsibility, leading two ML engineers through multimodal pre-training and staged text/image alignment for a visual-grounding model using a curated 1.1M-image dataset.
- Built distributed, multi-GPU training and serving infrastructure with AWS Batch, TensorRT, and Triton Inference Server, improving inference throughput by 5x over pure PyTorch serving.
- Built LoRA, Adapter, and LST post-training pipelines and delivered 1,100+ models and 60+ customer endpoints in one year.
- Built interactive segmentation tools using RepViT-SAM, FocalClick, SAM, and SAM2, increasing segmentation labeling speed by 25x.
- Reduced GPU memory usage by 65.6% and training time by 44.3% using parameter- and memory-efficient training methods.
- Won 2nd place in IOD and 4th place in FSOD challenges at CVPR 2025.
2021/02-2021/08: Machine Learning Engineer Intern
- Kakao Enterprise
- Fixed AutoGluon NeuralNetFastAI scaling issue.
- Developed a Flask-based training and inference AutoML framework with a simple front-end.
2019/02-2020/06: Co. Team Island CTO
- Team Island
- Built ZZAZZ (째즈), a mobile video-editing application that applied customizable motion effects to people using detection/segmentation, 3D mapping, and tracking.
- Built lightweight CNN models for mobile applications and on-device inference.
2018/08-12: Undergraduate Researcher
- UIUC Undergraduate Research Program
- Improved direction-of-arrival estimation with the MUSIC algorithm using irregular microphone arrays.
- Generated binaural sounds through a software-based audio pipeline.
- Supervisor: Professor Romit Roy Choudhury
2018/06-08: Machine Learning Engineer
- Seerslab Intern
- Developed face landmark detection using Haar cascades, HOG features, and machine learning methods.
- Built a GUI tool for annotating face coordinates.
- Developed CMS login functionality with JWT-based authentication.
2017/03-06: Undergraduate Researcher
- POSTECH Undergraduate Research Program
- Developed a non-invasive heart-rate measurement device.
- Collaborated on a smart-watch FPGA module for a national research project.
- Supervisor: Professor Park Sung Min
2016/06-08: Research Intern
- ASAN Medical Center
- Designed simulations for an electrical surgical unit using electromagnetic field analysis tools.
- Supported development and experiments for an assistive device for knee-injured patients.
- Supervisor: Professor Choi Jae Soon

Skills

Research: PyTorch, Hugging Face Transformers, multimodal pre-training and post-training, LoRA, contrastive alignment, evaluation
AI systems: AWS Batch, distributed multi-GPU training, TensorRT, Triton Inference Server, MLflow, model conversion
Programming: Python, Kotlin, Git; large open-source codebase development, testing, and documentation
Deployment: mobile ML, lightweight CNNs, on-device inference, TensorFlow Lite
AI developer tools: Claude Code, Codex, opencode, LangChain, Langfuse
Hardware and tools: FPGA development with Xilinx, PyQt5
Languages: Korean (native), English (professional)

Publications

Terahertz Signal Analysis in Tissue of Gastroesophageal Reflux Disease

Choi Sang Bum. (2013). &Terahertz Signal Analysis in Tissue of Gastroesophageal Reflux Disease." OPTICAL SOCIETY OF KOREA. POSTER SESSION I(WP-VI3).

On Joint Transfer of Energy and Information: A Markov Decision Problem Formulation

Sangbum Choi. (2018). "On Joint Transfer of Energy and Information: A Markov Decision Problem Formulation."None.

Improved DOA estimation of MUSIC algorithm based on irregular microphone array

Sangbum Choi. (2018). "Improved DOA estimation of MUSIC algorithm based on irregular microphone array." None. .

MobileHumanPose: Toward Real-Time 3D Human Pose Estimation in Mobile Devices

Sangbum Choi, Seokeon Choi, and Changick Kim. (2021). "MobileHumanPose: Toward Real-Time 3D Human Pose Estimation in Mobile Devices." IEEE/CVF CVPR Workshops, 2328-2338.

ZERO: Multi-modal Prompt-based Visual Grounding

Sangbum Choi and Kyeongryeol Go. (2025). "ZERO: Multi-modal Prompt-based Visual Grounding." arXiv.

Talks

Teaching

Electric Circuit

Linear Algebra

Athlete Service and Leadership

Approximately one year of team leadership at SuperbAI, including leading two ML engineers on multimodal pre-training.
Co-founder and CTO, Team ISLAND, 2019/02-2020/06.
Led up to eight people simultaneously across leadership roles.
Open-source contributor to Hugging Face Transformers