• Competition Link
• GitHub Link
• Role: Team Leader
• Project: Correction of weather forecast errors caused by the distance between prediction and observation points
• Goal: Improve day-ahead (24-hour) weather prediction accuracy for industrial sites (solar power, gas turbines), enabling better operational safety and combustion stability
• Model: Ensemble of XGBoost (capturing structured patterns via GBDT) + MLP (capturing nonlinear & latent patterns)
• Achievement: Grand Prize (산업통상자원부 장관상) for excellence in data utilization and industrial impact

• Project: FinCoach - AI-powered financial coaching service (Credit card delinquency prediction & spending pattern improvement recommendation service based on user MyData)
• Role: DA & DE
• Duration: 2025.08 ~ 2025.10 | Team of 6 (FE 2, BE 2, Data 2)

Key Achievements:

• Realistic MyData generation pipeline (DE): Combined AI HUB financial synthetic data + Statistics Korea household survey data, reflecting age-specific spending patterns with individual preference-based diversity using the principle "individuals behave uniquely, groups follow statistics"
• Credit card delinquency prediction model (ML): HistGradientBoostingClassifier-based model with 46 engineered features (13 personal info + 33 monthly derived variables including early/mid/late-month spending ratios, credit/debit ratios, 7 major category amounts/ratios, average transaction size, weekend/night spending ratios, HHI, entropy, 3-month rolling, EOM prediction, etc.)
• Spending habit improvement recommendation system: Hybrid CF + CBF approach using cosine similarity to extract top 5% similar users, analyzing 4 comparative perspectives: (1) current month vs similar users (2) 3-month average vs similar users (3) current month vs last month (4) current month vs personal 3-month average
• FastAPI-based prediction/recommendation API: Developed endpoints for delinquency prediction and spending comparison recommendations with real-time MySQL integration

Tech Stack: Python, scikit-learn, pandas, NumPy, FastAPI, pydantic, MySQL

Acheivement: Excellence Award (우수상) from Samsung Electronics

• Competition Link
• Code Share
• Project: Predicted the average shipment delay over the next 30 minutes for AMR-based smart logistics warehouses (regression, metric: MAE)
• Approach: Built 545 EDA-driven features and a 5-model ensemble (XGBoost L1/Quantile, MLP, Dilated CNN, Bidirectional Transformer), fused via MAE-optimized weighted averaging
• Core idea — Scenario-level priors: Trained a separate LightGBM to predict each scenario's overall mean & max delay, then injected those predictions as leakage-free fold-aligned OOF features into the row-level models. This passed scenario-wide difficulty down as a prior — and became the single strongest feature-importance signal, the main driver of the final score.
• Robust validation: Stress-stratified GroupKFold (4-axis load score) to keep high-load scenarios balanced across folds while preventing scenario leakage
• Achievement: 3rd Place out of 607 Teams (solo participation)

• Competition Link
• Project: Predicted subway congestion levels using time-series weather and observational data
• Reasoning: Dataset had numerous categorical variables, outliers, and missing values → CatBoost was more effective than specialized time-series models in this case
• Model: CatBoost (robust for categorical features, missing values, and outliers)
• Achievement: Finalist & Honorable Mention

• Competition Link
• Role: Team Leader
• Project: Developed an object detection model to identify potholes using real-world road images
• Model: YOLOv8
• Achievement: 8th place (Top 3%)

• Competition Link
• GitHub Link
• Role: Team Leader
• Project: Advertisement Click-Through Rate (CTR) Prediction Model
• Goal: Predict user ad click probability using user demographics, ad attributes, domain features, and behavioral sequences
• Model Architecture: Ensemble of two approaches
  • Model 1: CatBoost (base) + Transformer (sequence processing)
  • Model 2: xDeepFM (base) + Attention mechanism (sequence processing)
• Key Features:
  • Sequence modeling with Transformer encoder for user behavior patterns
  • Deep feature interaction modeling with xDeepFM's Compressed Interaction Network (CIN)
  • Attention-based pooling for sequence aggregation
  • Hybrid ensemble combining tree-based and deep learning approaches
• Achievement: Advanced to Final Round through effective fusion of tabular and sequential modeling
• Tech Stack: Python, PyTorch, CatBoost, pandas, NumPy, scikit-learn

• Competition Link
• Role: Team Leader
• Project: Classified used car types through image-based computer vision modeling
• Model: Ensemble of Swin Transformer and ConvNeXt
• Achievement: Top 13%

• Role: Team Leader
• Project: Designed obstacle-avoidance and high-speed driving logic in a virtual autonomous driving environment
• Achievement: 2nd place

• Project: Built a 6-month quest-based savings product where users earn EXP by completing Life/Growth/Surprise quests, level up for preferential interest rates, and contribute bonus interest to school-level donation pools
• My Role:
  • Developed and maintained the recommendation system end-to-end (hybrid CF+CBF with cold-start fallback, FastAPI + SQLAlchemy, interaction logging)
  • Contributed to the app’s front-end by handling design-oriented tasks in React Native (screens, components, styling)
• Achievement: Final Round (On-site)

• Study: Endomorphism of the 4-torsion group of elliptic curves
• Context: Participated as an undergraduate in a graduate-level academic conference
• Achievement: Outstanding Presentation Award (우수발표상)