Open to ML / research-engineer roles & research collaborations
AI/ML Engineer · Seoul, Korea
I build AI systems that hold up in production. As the primary engineer on four live government platforms at Gractor, I build the RAG pipelines, tool-calling agents, and evaluation harnesses they run on. Alongside that I publish first-author segmentation research: four papers on learning dense prediction from almost no labels. I care about the gap between what works in a paper and what a city can actually depend on.
The systems I've actually shipped at government scale — plus open-source tools and research side-projects.
A production RAG system for a city's water utility — hybrid retrieval + reranking over 2.7M+ meter readings, tuned to 98.8% accuracy on our eval set. It reasons over a TypeDB knowledge graph that's flagged thousands of leaks and contamination risks.
A tool-calling agent for the Anyang water grid (26,000+ meters). I rebuilt a sprawling LangGraph stack far simpler — a third less code, ~12× faster, and 100% on a red-team eval harness — reading the live system read-only so it can't break it.
A live 3D digital twin of a 215 km motorway (Zaragoza–Barcelona) — a ₩5.1B Korea–Spain program, Grand Prize at the 2024 City Design Awards. I took it from demo to certifiable on PostgreSQL/PostGIS, streaming real-time traffic onto a Three.js map.
Edge AI for a city of 460,000 — YOLOv5 on-device on smart-pole cameras, with MQTT and serial control to the hardware. It's offline-first: survives cloud outages and reboots a device if it goes quiet.
Read, fill & edit Korean HWP (Hancom Office) documents in Python — text extraction for LLM/RAG, programmatic form-filling, and corruption-free binary rewrite.
Stealthy backdoor triggers in CLIP-style models via dual-embedding alignment in the joint visual-textual representation space.
Targeted perturbations for multi-label classifiers that exploit semantic label co-occurrence dependencies.
Systematic evaluation of weak-to-strong consistency (UniMatch) and self-training (ST++) — the foundation for CW-BASS and FARCLUSS.
Depth estimation and point-cloud generation for autonomous driving, with an improved depth-to-pointcloud pipeline on KITTI.
Why declared context length overstates the effective one — analyzing positional-encoding bias and the STRING shifting method across Llama, Mistral & Qwen.
AI-powered disease identification with severity scoring, structured learning modules, community data contributions, and downloadable datasets.
20+ subjects, 70+ lessons, progress tracking with study streaks, a learning-analytics dashboard, full auth, and a responsive UI.
Four first-author papers, four different mechanisms — boundary-aware confidence weighting, fuzzy pseudo-labels with class rebalancing, contrastive pixel memory, and calibration-gated selection. CW-BASS & FARCLUSS were #3 and #2 on the Papers with Code SSSS leaderboards at release, in the ResNet-50/101 era; PixCon and CW-BASS v2 carry the line onto DINOv2 features.
Semi-supervised learningSemantic segmentationPseudo-labeling & consistencyCalibration & uncertaintyContrastive representation learningLong-tailed / class imbalanceBoundary-aware learningFuzzy & soft labelsFoundation models (DINOv2, ViT)Selective predictionLearning-theoretic analysisEvaluation methodology
orcid.org/0009-0004-7340-1873
Foundation-model teachers change the regime: DINOv2 confidence saturates, so the adaptive thresholding that helped weak teachers can hurt strong ones. A one-pass reliability gate measures whether the teacher's confident set is as reliable as the confidence it demands, then picks strict filtering or an adaptive floor — and picks correctly on all six DINOv2 teachers blind. Places 2nd among DINOv2 methods on Pascal VOC and Cityscapes, and beats the strict baseline by +1.5 mIoU on ADE20K, where the confident set is unreliable.
A clean-positive pixel memory bank that admits only labeled pixels, keeping contrastive positives free of pseudo-label noise on DINOv2-scale features. Ranks #2 in semi-supervised segmentation, matching a strong UniMatch V2 baseline at lower cost.
Fuzzy top-K pseudo-labeling, entropy-normalized uncertainty weighting, per-batch class rebalancing, and prototype contrastive regularization. #2 on the Papers with Code SSSS leaderboards at release (2025): 78.8% / 78.2% mIoU on Cityscapes & Pascal VOC, ResNet-101.
Confidence-weighted cross-entropy with Sobel boundary regularization and bounded-sigmoid dynamic thresholding. #3 on the Papers with Code SSSS leaderboards at release (2025): 77.15% mIoU on Pascal VOC 2012, ResNet-50.
Gractor · Smart-city AI · Seoul, Korea
The main engineer on four live government platforms. Most of my work is the RAG and agent systems a city's water utility runs on — retrieval at 98.8% accuracy over millions of live readings, and a tool-calling agent I rebuilt to a perfect eval score with a third less code. I also took a ₩5.1B Korea–Spain digital twin from demo to certifiable and shipped YOLOv5 edge inference on smart-city poles.
Korea University · Pattern Recognition & ML Lab
Advised by Prof. Seong-Whan Lee (IEEE Fellow). Two first-author segmentation papers, #3 and #2 on the Papers with Code leaderboards at release (2025); ~10K LOC of PyTorch multi-GPU training infrastructure; Korean patent filed (autonomous-driving perception).
GliT (GLITEC) · EdTech
Built offline-first mobile learning products reaching 500+ students and 80,000+ learning sessions, and led STEM-education initiatives.
Korea University · 2023 – 2026
Global Korea Scholarship (sole Zimbabwe awardee) · BK21 Research Fellowship · Advised by Prof. Seong-Whan Lee (IEEE Fellow).
Recognized at the Korean National Assembly
