Hello, I'm
ML Engineer & Data Scientist
I build ML systems judged by what they move — a loss ratio, an approval rate, a completion curve. First-author publication in bioinformatics, recommendation systems used by millions, and now decision intelligence at DEIN.
Based in Seoul — happiest shipping from anywhere with a good view.
The part of ML that's easiest to skip is the last mile: making a model actually change a decision. That's the only version I find worth doing. Right now that's decision intelligence at DEIN — building ML for insurance and loan judgments measured by loss ratio and approval rate, not accuracy on a held-out set.
I've worked across applied ML — recommendation and knowledge tracing used by millions, computer vision on real construction sites, and peptide-identification research I published as first author. The throughline isn't one technique; it's carrying a problem from "no one's framed this yet" to something that ships and holds up in the real world.
The domain keeps changing — bioinformatics, edtech, fintech — but the question doesn't: how does something that works in a notebook move something real in the world?
Trajectory
Not on the resume, but part of the picture — more about me →
Hanyang University — M.S. in Artificial Intelligence
Sep 2020 – Aug 2022
Advised by Prof. Eunok Paek. Thesis: NovoRank — Machine Learning Based Post-processing for Performance Improvement in De Novo Peptide Sequencing.
Inha University — B.S. in Civil Engineering & Software Convergence Engineering
Mar 2014 – Feb 2020
DEIN — Data Scientist
Aug 2026 – Present
Designing data schemas and KPI definitions for consumer-lending products at partner banks across Southeast Asia, and building an LLM agent that compiles natural-language specifications into executable data logic.
Ailys (now DEIN) — Data Scientist
Apr 2025 – Jul 2026
Built ML for insurance risk and loan approval decisions on the DEIN platform — judged by loss ratio and net interest margin, not accuracy alone.
TmaxEduAI — AI Engineer
Jan 2023 – Nov 2024
Built knowledge tracing and recommendation systems serving 2M+ learners — lifting course enrollment 5% and content completion 25% — shipped via Airflow pipelines. Awarded the Super Rookie Award, 2024 — the company's sole company-wide honor for a first-year employee.
Bioinformatics and Intelligent Systems Lab, Hanyang University — Graduate Research Assistant
Aug 2020 – Dec 2022
Designed ML/DL-based post-processing tools for reranking and rescoring, lifting average recall and precision by 4.5% and 4.6% across three benchmark de novo sequencing tools — published as first author in J. Proteome Res. (2025).
The Construction Systems Laboratory, Inha University — Undergraduate Researcher
Jul 2019 – Feb 2020
Developed image classification and object detection models to detect hazardous objects in indoor construction environments, raising CNN accuracy from 85.7% to 87.6% through targeted image augmentation.
NovoRank: Refinement for De Novo Peptide Sequencing Based on Spectral Clustering and Deep Learning
Jangho Seo, Seunghyuk Choi, Eunok Paek — J. Proteome Res. 2025, 24, 2, 903–910
Read paper →Talks & Presentations
Post-Processing Framework Using Deep Learning to Improve De Novo Peptide Sequencing Results
Jun 2022
Poster · 70th ASMS Conference on Mass Spectrometry and Allied Topics, Minneapolis, MN
Image Augmentation for Small Object Detection on Indoor Construction Site
2019
Oral presentation · Korea Institute of Construction Engineering and Management (KICEM), Goyang, Korea
Featured
Decision Intelligence · Insurance & Loan
At DEIN (formerly Ailys) I build ML judged by business KPIs, not accuracy on a held-out set — an insurance disease-risk model and loan reject inference that expanded approvals with no change in delinquency.
>93%
Disease-risk recall
+4.6%
Loan approvals
+$2.21M
Net interest margin
Bioinformatics · Deep Learning · Clustering
Conventional de novo peptide sequencing tools score spectra in isolation and frequently misidentify peptides. I built a two-step clustering and deep learning re-ranking pipeline that cross-checks candidates against similar spectra.
+4.6%
Avg. precision gain
+4.5%
Avg. recall gain
Across three state-of-the-art tools
More projects
Recommendation Systems · Knowledge Tracing
Course recommendations for a 2M+ user platform were generic and didn't reflect learner progress. I built a deep learning recommendation model on sequential interaction data and shipped it through an Airflow pipeline, lifting course enrollment (CVR) by 5% and content completion by 25%.
Computer Vision · Image Augmentation
Recognizing small hazard objects on construction sites is data-hungry, but labeled images are scarce. I studied which image-augmentation strategies actually help, lifting a CNN's accuracy from 85.7% to 87.6% — and found rotation hurts more than it helps for these objects.