I build AI systems that hold up in production — the unglamorous part where a demo becomes something a city actually runs on. I'm the primary engineer on four live government platforms at Gractor, a smart-city AI company: RAG over 2.7M+ sensor records at 98.8% accuracy, tool-calling agents, and the evaluation harnesses that keep them honest.
During my MSc in AI at Korea University's Pattern Recognition & Machine Learning Lab (advised by Prof. Seong-Whan Lee, IEEE Fellow) I published two first-author papers on semi-supervised semantic segmentation. I now continue that line independently — four papers, the two newest on foundation-model backbones.
My research sits where learning from limited labels meets dense prediction, and it moves across the stack: loss design and boundary-aware regularization, fuzzy and soft pseudo-labels, uncertainty quantification and model calibration, contrastive representation learning, long-tailed class rebalancing, theoretical analysis of estimators and gradients, and the multi-seed evaluation methodology that keeps such results honest — on CNN and foundation-model backbones alike.
| Year | Paper | Venue | Rank ‡ | Links |
|---|---|---|---|---|
| 2026 | CW-BASS v2 — Saturation-Aware Pseudo-Label Selection under Foundation-Model Teachers | Under review IEEE TPAMI |
2nd DINOv2 |
arXiv · Code · Project · Models |
| 2026 | PixCon — Clean-Positive Contrastive Learning for Foundation-Model SSSS | Under review WACV 2027 |
#2 | arXiv · Code · Project |
| 2026 | FARCLUSS — Fuzzy Adaptive Rebalancing & Contrastive Uncertainty Learning for SSSS | Neural Networks 2026 * Q1 · Top 10% IF |
#2 | arXiv · Code · Project |
| 2025 | CW-BASS — Confidence-Weighted Boundary-Aware Learning for SSSS | IEEE IJCNN 2025 † CORE A (2020) |
#3 | IEEE · arXiv · Code · Project |
Four papers, four different mechanisms: boundary-aware confidence weighting with dynamic thresholds (CW-BASS); fuzzy top-K pseudo-labels, entropy-based uncertainty weighting and adaptive class rebalancing (FARCLUSS); a contamination-free pixel memory bank with a supervised-InfoNCE gradient analysis (PixCon); held-out calibration and a reliability-gated selection rule (CW-BASS v2). The first two were #3 and #2 on the Papers with Code leaderboards at release (2025, ResNet era); the last two move the line onto DINOv2 backbones.
Semi-supervised learning · Semantic segmentation · Pseudo-labeling & consistency regularization · Model calibration & uncertainty quantification · Contrastive & metric representation learning · Class imbalance / long-tailed learning · Boundary & edge-aware learning · Fuzzy logic & soft labels · Foundation models & transfer (DINOv2, ViT vs CNN) · Selective prediction & reliability estimation · Learning-theoretic analysis (estimators, bounds, gradients) · Empirical evaluation methodology (multi-seed, variance, ablation protocol)
Formal classification (ACM CCS · IEEE EDICS)
ACM CCS — I.4.6 Segmentation · I.4.6.d Pixel classification · I.2.10 Vision and Scene Understanding · I.2.6.g Machine learning · I.2.6.c Connectionism and neural nets · I.5.2.a Classifier design and evaluation · D.2.4.h Statistical methods
IEEE EDICS — Computer Vision → Segmentation, grouping and shape analysis · Computer Vision → Self-, semi-, meta- and unsupervised learning
‡ Ranks are at release, not live standings — Papers with Code closed in July 2025. CW-BASS v2's 2nd is among DINOv2-backbone methods, behind UniMatch V2-B.
* Neural Networks (Elsevier) is the premier neural networks journal — Q1 and Top 10% by impact factor (JCR), and the archival journal of the world's three oldest neural-network societies: the International (INNS), European (ENNS) and Japanese (JNNS) Neural Network Societies.
† IEEE IJCNN is the premier international conference in neural networks, run by INNS with the IEEE Computational Intelligence Society; CORE A (2020).
| Project | What it does | Stack |
|---|---|---|
| hwpkit | Read, fill & edit Korean HWP (Hancom Office) docs in Python — text extraction for LLM/RAG, programmatic form-filling, and corruption-free binary rewrite | Python · OLE/CFB |
| Claude Usage Widget & Token Tracker | Live system-tray widget for Claude Code plan limits (5h/7d) + local token & cost analytics per project, model, and tool | Python · GTK · CLI |
Claude Code usage — I build with agentic coding daily.
Local Claude Code telemetry, snapshot updated Aug 2026.
AI / LLM Systems
-6AA6FF?style=flat-square){kind=icon}
ML / Computer Vision
Backend & Data
Frontend
Infra / DevOps / IoT
Languages
AI/ML Engineer · Gractor Co., Ltd. · Sept 2025 – present · Seoul
Primary engineer across four live government platforms. Built a RAG system over 2.7M+ IoT records at 98.8% eval accuracy, rebuilt a production agent from 94.9% → 100% (96/96) with ~30% less code and ~12x faster startup, shipped a multi-provider LLM router with circuit-breaker failover, and deployed YOLOv5 + OpenVINO edge inference on government smart poles.
Research Engineer (MSc) · Korea University, PRML Lab · Sept 2023 – Feb 2026 · Seoul
Advised by Prof. Seong-Whan Lee (IEEE Fellow). Two first-author segmentation papers there (#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).
AI Software Engineer · GliT (GLITEC), EdTech · Jan 2019 – Jan 2021 · Zimbabwe
Built two offline-first mobile learning products reaching 500+ students and 80,000+ learning sessions.
MSc in Artificial Intelligence · Korea University · 2023–2026 · GPA 3.78/4.0
Global Korea Scholarship (sole Zimbabwe awardee) · BK21 Research Fellowship · Advisor: Prof. Seong-Whan Lee
Awards — GINCON Global Award 2025 (Korean National Assembly)
Let's build something that ships.
Like this profile? Reuse the design with attribution to Ebenezer Tarubinga. See LICENSE.