Open to ML / research-engineer roles & research collaborations
Ebenezer Tarubinga · AI/ML Engineer · Seoul, Korea
Four first-author papers on semi-supervised segmentation — teaching a model to label every pixel from a handful of examples, without letting its own mistakes compound. I work the same problem in production: retrieval, agents and computer vision for smart-city infrastructure.
4
First-author papers
2
Peer-reviewed
#2
Global rank at release
1
Korean patent filed
01 — Building
Production systems, open-source tooling and research side-projects. The production work is described by what it does and what I built, not by whose data is in it — that part belongs to the client, so the repositories stay private.
A retrieval system for a municipal water utility: hybrid retrieval with reranking over operational meter data, reasoning across a knowledge graph so an operator gets an answer with its evidence attached. Scored against a held-out evaluation set, not impressions.
A tool-calling agent over a municipal water grid that reads the live system strictly read-only, so it cannot act on infrastructure. I rebuilt a sprawling agent stack substantially smaller and faster, with guardrails at every tool boundary and an evaluation harness that covers adversarial inputs.
A live 3D twin of motorway infrastructure for an international smart-mobility programme — real-time telemetry streamed onto a Three.js map over PostgreSQL/PostGIS. Most of my work was hardening a prototype into a build that could pass certification.
On-device detection running on municipal camera poles, with MQTT and serial control down to the hardware. Offline-first by design: it keeps working through a cloud outage and recovers a device that stops reporting.
Read, fill and edit Korean HWP (Hancom Office) documents in Python: text extraction for LLM and RAG pipelines, programmatic form-filling for government and university paperwork, and a binary rewrite that does not corrupt the file.
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 rather than treating each label independently.
A systematic evaluation of weak-to-strong consistency (UniMatch) and self-training (ST++) — the groundwork the four papers were built on.
Depth estimation and point-cloud generation for autonomous driving, on an improved depth-to-pointcloud pipeline over KITTI.
Why a declared context length overstates the effective one — positional-encoding bias and the STRING shifting method, measured across Llama, Mistral and Qwen.
Disease identification with severity scoring for smallholder farmers, plus structured learning modules, community data contributions and downloadable datasets.
O Level and A Level revision for African high-school students: 20+ subjects, 70+ lessons, study streaks, a learning-analytics dashboard and full auth.
02 — Research
Four first-author papers, four mechanisms for one problem — boundary-aware confidence weighting, fuzzy pseudo-labels with class rebalancing, contrastive pixel memory, and calibration-gated selection. CW-BASS and 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.
01
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.
Picks correctly on all six DINOv2 teachers, blind · 2nd among DINOv2 methods on Pascal VOC and Cityscapes · +1.5 mIoU on ADE20K, where the confident set is unreliable
02
A clean-positive pixel memory bank that admits only labeled pixels, keeping contrastive positives free of pseudo-label noise on DINOv2-scale features. One branch over a consistency backbone, with nothing added at inference time.
Ranks #2 in semi-supervised segmentation, matching a strong UniMatch V2 baseline at lower cost
03
Fuzzy top-K pseudo-labeling, entropy-normalized uncertainty weighting, per-batch class rebalancing and prototype contrastive regularization — generalization without discarding the pixels the model is least sure about.
#2 on the Papers with Code SSSS leaderboards at release (2025) · 78.8% / 78.2% mIoU on Cityscapes & Pascal VOC, ResNet-101
04
Confidence-weighted cross-entropy with Sobel boundary regularization and bounded-sigmoid dynamic thresholding, aimed at the two failures that compound each other: boundary blur and confirmation bias.
#3 on the Papers with Code SSSS leaderboards at release (2025) · 77.15% mIoU on Pascal VOC 2012, ResNet-50 · 65.9% on Cityscapes from 100 labeled images
03 — Path
Gractor · Smart-city AI
Sept 2025 — present · Seoul, Korea
Retrieval, agents and computer vision for smart-city platforms — edge to cloud. I build the RAG and agent systems, the guardrails around them, and the evaluation harnesses that keep them honest once real operators are using them.
Korea University · Pattern Recognition & ML Lab
Aug 2023 — Feb 2026 · Seoul, Korea
Semi-supervised segmentation under Prof. Seong-Whan Lee (IEEE Fellow): four first-author papers, roughly 10K lines of multi-GPU PyTorch training infrastructure, and a Korean patent filing for autonomous-driving perception.
GliT · EdTech
Jan 2019 — Jan 2021 · Hybrid
Offline-first mobile learning products reaching 500+ students and 80,000+ learning sessions, ten-plus full-cycle product builds, and STEM programmes run with partner schools.
Korea University · 2023 – 2026
Computer vision, advised by Prof. Seong-Whan Lee (IEEE Fellow).
Republic of Korea · application filed
Real-time object detection method, recording medium and device performing the same — filed out of the autonomous-driving perception work.
04 — Toolkit
RAG (hybrid retrieval), agents & tool-calling, Model Context Protocol, evaluation harnesses, guardrails, LangGraph, OpenAI, Anthropic Claude, HyperCLOVA X, Hugging Face
PyTorch, TensorFlow, OpenCV, scikit-learn, NumPy, Pandas, SciPy, YOLOv5, OpenVINO, ONNX, CUDA, Weights & Biases
FastAPI, Flask, Node.js, Express, SQLAlchemy, Prisma, knowledge graphs, PostgreSQL + PostGIS, MongoDB, Redis, OpenSearch
React 19, TypeScript, Next.js, Three.js, deck.gl, MapLibre, ECharts, Tailwind, Radix UI, Framer Motion, Zustand, Vite
Docker, Kubernetes, AWS, Nginx, systemd, Prometheus, Grafana, Mosquitto MQTT, ModBus RTU, Vercel, Cloudflare, GitHub Actions
Python, TypeScript, JavaScript, C++, C#, Java, SQL, Bash
05 — Credentials
Applied AI Professional Certificate
IBM
Modern Robotics Specialization
Northwestern University
Semantic Segmentation with SageMaker
Amazon Web Services
ML Pipelines with Azure ML Studio
Microsoft
Open to ML and research-engineer roles, and to research collaborations.
