End-to-end ownership
I own the full path — data extraction and feature engineering, training and evaluation, then serving, monitoring, and retraining. No hand-off gaps where models quietly break.
$ Data scientist · Machine learning · Seoul
I turn data into decisions
I build machine learning systems that ship — from feature pipelines to production models — and I measure their impact in dollars, latency, and error rate. Explore the case studies below.
How I work
From messy data to models people trust
I care less about the fanciest architecture and more about whether the model holds up in production and moves a real metric. Here's how I approach the work.
Rigorous evaluation
Offline metrics are table stakes. I back every model with holdouts, causal experiments, and calibration checks so a lift is real and not just a lucky slice of the data.
Impact you can measure
Every project below states the outcome in the units the business cares about — churn points, forecast error, response time, revenue — not just the algorithm I reached for.
Selected work
Projects
Production machine learning across churn, forecasting, NLP, experimentation, fraud, and recommendations. Click any card for the role, methods, and measured impact.
FeaturedSubscriber Churn Prediction Engine
2024A gradient-boosted churn model that scores every subscriber nightly and feeds a retention playbook.
PythonXGBoostSparkSQL
View case study 
FeaturedRetail Demand Forecasting Platform
2024A hierarchical forecasting system that predicts SKU-level demand across 900 stores two weeks out.
PythonPyTorchDartsPandas
View case study 
Support Ticket Triage with NLP
2023A transformer-based classifier that routes and prioritises inbound support tickets in real time.
PythonPyTorchHugging FaceFastAPI
View case study 
Experimentation & Causal Inference Platform
2023A statistical engine that powers trustworthy A/B tests and untangles cause from correlation.
PythonRstatsmodelsSQL
View case study 
Real-Time Fraud Detection
2022A streaming anomaly-detection model that scores transactions for fraud within 50 milliseconds.
Pythonscikit-learnKafkaFlink
View case study 
Personalised Recommendation System
2022A two-tower recommender that powers homepage and email personalisation for millions of users.
PythonTensorFlowFAISSSpark
View case study Have a hard data problem?
Whether it's a model that stalled in production or a question your dashboards can't answer, I'd love to hear about it. Let's talk.