Jin-Ho Ju

github.com/Wlsghdh wlsghdh.github.io edu-data.tistory.com

Education

Suwon University — B.S. in Data Science

Mar 2021 — Feb 2027 (Expected)

GPA 4.33 / 4.5

Sungil High School

Mar 2019 — Feb 2021

Publications

Why Deep ResNets Train Successfully: Self-Selection of Effective Depth Enabled by Skip Connections 1st Author

arXiv · BMVC, under review (2026)

Jin-Ho Ju, Hong-Ryul Ahn

Proposed Learnable Residual Scaling (LRS) — a learnable scalar α per residual block in the form y = α·F(x) + (1−α)·x, where α itself directly measures per-block utilization. Showed that deep ResNets self-select an effective depth far smaller than their nominal one (only 5–6 active blocks out of 200). Validated a depth–α scaling law on CIFAR-10/100 (ResNet-50/101/152/200) and ImageNet (ResNet-50/101); confirmed zero accuracy loss without retraining when pruning blocks with α < 0.03 on ResNet-200/CIFAR-100.

github.com/Wlsghdh/Learnable_Residual_Scaling

Effects of Generative-AI Augmentation for Small-Sample Industrial Defect Detection 1st Author

KIISE 2026

Jin-Ho Ju, Dae-Yoon Lim, Jin-Woo Yang, Hong-Ryul Ahn

Tackled the structural data scarcity of industrial defect detection (~20 samples per class). Demonstrated that semantic-level diversity matters more than quantity: with Mask R-CNN, traditional augmentation hurt mAP (−1.76), while Gemini-2.0-based generative augmentation improved it (+1.90) on the same volume. Found an 8× sweet spot when combining both (+5.0 mAP), consistent across 2-stage CNN, 1-stage CNN, and Transformer detector families.

github.com/Wlsghdh/VISION-Instance-Seg

An Integrated Preprocessing Pipeline for Model Performance Comparison on a Multimodal Gas Sensor Dataset 2nd Author

KIISE 2025

Y. Maeng, Jin-Ho Ju, J. Yoon, W. Jung, Hong-Ryul Ahn

Designed a 4-stage standardized pipeline (anchor detection → outlier removal → min-max normalization → reproducible 20-rep CV split) for the MultimodalGasData benchmark. Under this fair-comparison protocol, benchmarked 8 models and revealed a 22.2 %p accuracy gap between best (MT-Fusion E+D+C, 97.8 %) and worst (Random Forest, 75.6 %) — proving that fusion strategy outweighs model depth.

github.com/Ahn-Laboratory/Gas-Leakage-Detection

Selected Projects

CarNeRF — AI-powered used-car dealer platform

Dec 2025 — Present

Personal / Team Project · In active competition

End-to-end web & mobile platform where a 1-minute seller video triggers an automated pipeline: FastGS-based 3D reconstruction (CVPR 2026) → YOLOv8 exterior-damage detection → LightGBM price prediction → fake-listing classification → LLM-driven natural-language search. Built the entire 3D pipeline from scratch (frames → COLMAP → rembg + depth → Gaussian Splatting → SPLAT export → web viewer), reaching PSNR 31.88 on vanilla 3DGS and a ~15× speed-up after migrating to FastGS (60K → 30K iter ≈ 2–3 min on A100). Resolved pycolmap segfaults via OPENBLAS thread tuning and shipped a React Native (Expo) beta client.

FastGS · COLMAP · SAM · YOLOv8 · LightGBM · FastAPI · SQLAlchemy · RAG · React Native

github.com/Wlsghdh/CarNeRF · lifeai.suwon.ac.kr:5199

ETF with AI — LambdaRank-driven ETF / stock trading system

Feb 2026 — Present

Personal / Team Project · In active competition

Integrated trading system combining a LightGBM LambdaRank (Learning-to-Rank) model over 85 features (technical indicators + macroeconomic signals + z-scores), an automated TradingView scraper (Playwright + SSH-tunneled MySQL), and a Next.js monitoring dashboard. Connected to the KIS API for real-money execution with cron-driven daily prediction and monthly retraining. Currently scaled to 1,000 tickers with a 3×3 multi-AI fusion grid (Technical / Fundamental / Market).

LightGBM LambdaRank · FastAPI · Next.js · Playwright · MySQL · Docker Compose · KIS API

github.com/Wlsghdh/etf-trading-projects · ahnbi2.suwon.ac.kr/trading

JUMP AI 2025 — 3rd Drug Discovery Competition (solo) Top 4% · Rank 20 / 1,134

Jul — Aug 2025

DACON

Built a MAP3K5 IC50 activity predictor from SMILES inputs for drug-candidate screening. Engineered 140-dim features (RDKit descriptors + PCA-compressed Morgan ECFP4 fingerprints), tuned 5 base learners (LightGBM / XGBoost / RF / Extra Trees / MLP) with Optuna (30 trials × 5-fold CV), then combined them via a two-track ensemble blend (SLSQP-weighted + quantile-matching).

RDKit · LightGBM · XGBoost · Optuna · scikit-learn

github.com/Wlsghdh/Jump-AI-2025

Personal Color Diagnosis System Hwaseong City AI Forum, Invited Talk

Jun 2024 — Mar 2025

Personal / Team Project

Interactive single-booth web experience that diagnoses a user's personal color from a face photo and plays matching video content. A curiosity-driven study of whether CNNs — built for pattern recognition — could perform well on a pure color-composition task. Pipeline: collected ~50K celebrity reference images, applied white-balancing, segmented skin tone via OpenCV (eyes / nose / mouth removal), then routed the model output through an OpenAI prompt to a curated video / sound booth in real time.

PyTorch · OpenCV · OpenAI API · AWS · Docker · React

github.com/Woochang4862/personal-color-app

Experience

Military Service — Republic of Korea Marine Corps

Jan 2022 — Jul 2023

Completed mandatory military service.

Autonomous Driving Team Intern

Aug 2024 — Feb 2025

AIMMO Inc.

Cleaned vision-fail data, performed Labellerr-based labeling QA, collected training data, and analyzed sensor graphs from detection failures.

Undergraduate Research Assistant

Jul 2025 — Present

Suwon University, Ahn Laboratory

Computer-vision research that led to the first- and co-authored publications listed above. Built applied React / Node.js / MongoDB prototypes alongside the research work.

Research Intern — Hadd Science

Sep 2025 — Dec 2025

Sungkyunkwan University (SKKU)

Managed the company website, wrote articles, and produced 3D work in AutoCAD.

Teaching Assistant — DSML Vibe Coding

Feb 2026 — Present

Suwon University, DSML

Mentoring undergraduate students.

Awards & Honors

Academic Excellence Scholarship — Suwon University · Merit-based, awarded 5 semesters (Fall 2021, Fall 2024 – Spring 2026)

2021 — 2026
Encouragement Award — Korean Institute of Information Scientists and Engineers (KIISE)

Dec 2025
Top 4% · Rank 20 / 1,134 — JUMP AI 2025 Drug Discovery Competition, DACON

Aug 2025
1st Place — Stock Chart Data Collection #2, Suwon University DSML

Nov 2025
1st Place — Stock Chart Data Collection #1, Suwon University DSML

Oct 2025
Excellence Award — AI/Develops, Suwon University DSML

Aug 2024

Technical Skills

Deep Learning: PyTorch, PyTorch Lightning, LLM fine-tuning (LoRA / PEFT), DDP, Optuna, Weights & Biases
Computer Vision: YOLO, detectron2, mmdetection, OpenCV, Gaussian Splatting, COLMAP, SAM
NLP & LLM: HuggingFace, RAG, prompt engineering
Engineering: FastAPI, SQLAlchemy, Next.js, Docker, MySQL, Selenium, Playwright
Workflow: Git (PR / branch / issue / commit), Linux, AWS

Completed Programs

Google Data Analytics Professional Certificate

Feb 2025

Coursera × Google

Digital Healthcare (Beginner)

Nov — Dec 2025

GUIP Biohealth Platform

Why Me

Curiosity-driven research

I refuse to stop at shallow understanding. When a question lands, I design small experiments to chase it down myself — and the failures along the way have always been the most valuable part of the process.

Root-cause problem solving

With a habit of startup-minded tinkering, I rarely walk past an inconvenience. I ask why repeatedly until I'm looking at the underlying mechanism — whether it's an AI model or a paper I'm reading for the first time.

Growth as a team

I have led most of my collaborative projects. What I value most is matching each member's strengths to the right role — real synergy comes from communication, not from a sum of skills.