Projects
Research projects, production systems I built in industry, lab work during graduate school, and the coursework and apps that came before. Papers are on the publications page.
Research
Answer-position bias in MoE vision–language models
Independent project · Jan 2026 – present
Training-free diagnosis and repair of fine-grained discrimination failures in eight open MoE VLMs. A zero-cost diagnostic and a "calibration law" for training-free interventions.
Second-pass document-hierarchy repair
Laputa · 2026
Repairing the hierarchy PDF extractors get wrong: an end-to-end evaluation on three public benchmarks and Korean textbooks, a self-supervised pointer model trained without human labels, and LLM/VLM baselines.
Diffusion-based demosaicing and compression
Ph.D. thesis · 2021 – 2024
Image demosaicing and compression with a denoising diffusion null-space model; a learned color-filter-array pattern and a CFA optimization scheme for jointly optimal image quality and compression rate.
GAN super-resolution with feature-map adversarial loss
M.Sc. thesis · 2019 – 2021
A training scheme that applies adversarial loss on the feature maps of the generated high-resolution image; outperformed the state-of-the-art CVPR method on a real-world benchmark at the time.
Industry systems
Production work at STANS (AWAS-Insight, intelligent CCTV) and Laputa (Unibook, education technology).
Unibook: multimodal textbook Q&A system
Laputa · Dec 2025 – present
The AI Q&A feature of Unibook, an e-book platform for Korean textbooks: LLM-based RAG and agentic workflow around a 20B+ open-weight multimodal MoE model. Designed the agentic workflow, fixed and tuned the retrieval pipeline, redesigned its database, chose and validated the production model by measurement, and own the cloud-GPU deployment stack (RunPod, custom Docker images, EC2).
AWAS-Insight: crowd-density detection for Seoul metro
STANS · 2024 – 2025
Real-time people counting to prevent overcrowding; fine-tuned a crowd-counting model on company data, built the auto-labeling pipeline and annotation guidelines. Deployed at Seoul subway stations and other public areas.
AWAS-Insight: object detection
STANS · 2024 – 2025
Real-time object detectors fine-tuned on company datasets including LiDAR-camera data, quantized for real-time inference; deployed in the same Seoul metro systems. Also rule-based intrusion and loitering detection.
AWAS-Insight: school violence detection
STANS · 2025
Surveyed action-recognition models for deployability, then reimplemented an action-recognition network and an optical-flow/frame-difference model in PyTorch, integrated them with a real-time object detector, quantized, and prepared the system for KISA (Korea Internet & Security Agency) school-safety certification evaluation.
Construction-site real-time translation
STANS · 2025
Early-stage R&D for translating Korean construction jargon across multiple Asian languages: speech-to-text, neural machine translation, TTS, and LLM fine-tuning with RAG. My first LLM development work.
Graduate-school lab projects
Work at the Image Processing Lab, Dongseo University, beyond my thesis research.
DeepView multi-plane view synthesis for ETRI
Commissioned project · Dongseo
Reviewed and reimplemented DeepView: View synthesis with learned gradient descent for ETRI: learned gradient descent and multi-plane images for depth-based view synthesis, with reports and presentations to the institute.
Blood-vessel segmentation
Lab project
U-Net segmentation on a real medical image dataset, with data augmentation to compensate for the small amount of labeled data.
Crowd counting with CSRNet
Lab project
Supporting a labmate implementation and experiments with CSRNet for density-map crowd counting.
Class projects
Coursework during my M.Sc. and Ph.D. which include paper reviews, implementations, and write-ups, each came with a project paper and presentation I'm still fond of. Some are available in GitHub, I'm thinking of adding the PPTs here in the future.
Deep Q-Network on Sonic the Hedgehog 2
Reinforcement learning
DQN agent trained to play Sonic 2; I was later invited back to demonstrate it to the following semester's class.
Fooling a person re-identification model
Computer vision
Adversarial masking of face regions to study how re-ID models fail — an early brush with the fine-grained robustness questions I work on now.
Deep learning for efficient network routing
Computer networks
Applying deep learning models to routing decisions in computer networks.
Genetic algorithm for the travelling salesman problem
Optimization
Experiments with a proposed genetic-algorithm variant on TSP instances.
Curve interpolation from scratch
Computer graphics
Lagrange, Bézier, and cubic Hermite curve implementations.
Apps & web
Software I built before and around graduate school.
Physics learning app for high-school students
Volunteer program · Indonesia
An interactive physics learning app built as part of a social-service program to improve education quality in Indonesia.
Android chat application
PT. Mat Ali Teknologi · 2019
Developed an Android chat application.
Mobile quality-control system for an aluminium company
Freelance
A lightweight hybrid mobile app (Android and iOS) for aluminium storage and quality-control management.
Multi-branch procurement system for a property agency
Freelance
Web system for managing real-estate property and agent data across branches.
Committee management system
Bachelor thesis · University of Surabaya
Web app for the Department of Informatics' event committees: task tracking, team performance evaluation, and document management.