Dohyeop Lim

Dohyeop Lim

AI researcher & developer

Ask anything about me

Research

I work on data generation, model development, and evaluation for computer vision and multimodal models.

Publications

E-ACT, Enforcing Arithmetic Constraints in CTC Trellises for Structured Identifier Recognition

Dohyeop Lim, MinKi Jeong, Jongyoul Park (Under review)

A factorized CTC decoder that enforces identifier length, allowed characters, and checksums. Works with existing recognizers without retraining.

Defined the problem, implemented checksum tracking through the CTC trellis, and evaluated recognition accuracy and error propagation.

GreedyE-ACT admissible path

Standard OCR

Read text, then check the rules.

Validation can reject the reading, but cannot revise it.

E-ACT

Apply the rules during decoding.

Select a reading that satisfies the identifier rules.

Imagine the Structure Before You Speak: Structural Imagination in Latent Space for Table Recognition

MinKi Jeong, Dohyeop Lim, Jongyoul Park (Under review)

R&D participation

산업AI용데이터전처리자동화기술개발

Industrial OCR research with KETI & KOSPO.

산업AI용데이터전처리자동화기술개발

OCR preprocessing

Developed an integrated module for OCR-focused image preprocessing and compared preprocessing performance across 3,194 industrial MPSC images.

OCR preprocessing performance comparison on industrial MPSC images

E-ACT

Building on the OCR-focused preprocessing work from the first project year, developed E-ACT, a general-purpose module that improves identifier OCR decoding without additional training.

E-ACT decoding architecture
Qualitative examples of identifier recognition with E-ACT

Kraftbox

Working with KOSPO (Korea Southern Power), identified the difficulty of quickly and consistently checking handwritten safety work permits for missing required fields and entry errors. Developed Kraftbox, a synthetic document generation pipeline, to support models and pipelines for reviewing these permits.

Kraftbox synthetic document generation for safety work permits

컴퓨팅자원집중형인공지능응용기술개발

Vision-language-action model research with IITP.

컴퓨팅자원집중형인공지능응용기술개발

분산 GPU 기반 비전-언어-행동 모델 확장 핵심 기술 개발

Participating researcher

Assist with requirements and technical development for an education-focused vision-language-action (VLA) system.

Support GPU resource allocation and shared research environments with Docker.

문화서비스확산형기술개발

Multimodal content creation with ETRI.

문화서비스확산형기술개발

시니어의 콘텐츠 제작 접근성 향상을 위한 생성형 AI 기반 콘텐츠 창·저작 플랫폼 기술 개발

Participating researcher

Reviewed software, managed issues, and conducted QA in preparation for integration testing.

Selected projects

Additional experience

Visiting student

Technische Hochschule Ulm, Germany

January to February 2026

Developed DocFusionX to address the lab’s document retrieval needs.

DocFusionX

Community

LIKELION University SeoulTech Chapter

Vice President, 14th cohort. Member, 13th cohort.

2025 to present

Coordinate a 30-member chapter. Organized four technical sessions, invited talks, and hackathons.

Website

Google Developer Groups on Campus

Core team member, 5th and 6th cohorts.

2025 to present

Organized 13 technical sessions in the 5th cohort.

GitHub

Background

Seoul National University of Science and Technology

B.Eng. in Applied Artificial Intelligence

2024 to present

GPA 4.34 / 4.5, currently ranked first in the department. Academic Excellence Scholarship, full tuition for five consecutive semesters.

Skills

Programming
Python, Swift, TypeScript
Research & machine learning
PyTorch, OpenCV
Applications
SwiftUI, Core ML, Next.js
Tools
Linux, Docker, Git, vLLM, LaTeX, Figma
Spoken languages
Korean (native), English (TOEFL iBT 100/120)