Business × Computer Science · Korea University · Seoul (KST)
I work on the seam between analytics and decisions — cleaning messy data, finding the question underneath it, and shipping something a non-analyst can act on. Most of my work lands as a written report rather than a model file.
Currently focused on NLP over user reviews, demand forecasting, and making analysis reproducible enough that someone else can rerun it.
Portfolio · Writing · LinkedIn · Contact
| Project | What it does | |
|---|---|---|
| LCC Review Sentiment Cluster | Prioritized high-intensity, low-cost airline complaint fixes. | Report |
| Steam Review | Separated loud complaints from complaints that actually drive churn. | Report |
| ConflictDynamics | Co-built an LLM multi-agent office simulator where memory and shifting relationships drive conflict, with a Phaser replay viewer. | Collab |
More projects
- tinyLLM — Pretrained a 29.6M-parameter chat model from scratch in 3 hours on one RTX 3060 Ti.
- Toronto Airbnb Pricing — Compared what listing features are worth in Toronto's short-term and monthly Airbnb markets. (Report)
- Transformer from Scratch — Built a transformer from torch primitives, tested by the properties the paper relies on.
- deeptool — A thin helper library for handling PyTorch models object-oriented-style in Jupyter. (Docs)
- Seoul Bike Analysis — Mapped station-level demand and supply gaps across Seoul.
- argos — A local-first macOS desk where chat messages become tasks, events and notes for AI agents.
All projects → sciencemj.github.io/projects.html
- 2026-11 · What Drives Airbnb Prices in Toronto? — Report
- 2026-08 · Attention Is All You Need 논문 요약 — Post (KO)
- 2026-08 · Batch Normalization — Post (KO)
NLP · sentiment · clustering · time-series forecasting · data visualization · deep learning