CS major working at the intersection of AI agents and the web.
I don't just use AI — I build the guardrails and products that make it dependable.
🛡️ Currently building grounded — a deterministic guardrail that makes AI coding agents prove it before they act.
- AI Agent Reliability — guardrails and hooks that enforce evidence-based agent behavior
- LLM & Prompt Engineering — building with large language models, beyond the happy path
- RAG (Retrieval-Augmented Generation) — document-grounded AI chatbots
- LangChain / LlamaIndex — AI application pipelines
- React / Next.js — server-side rendering and full-stack web apps
- TypeScript — type-safe frontend development
- AI-powered Web Apps — web services built on LLM APIs
Make AI coding agents prove it before they act.
A deterministic guardrail that blocks ungrounded agent actions — editing files it never read (including sed -i / tee shell bypasses), installing hallucinated packages, citing dead links. No LLM in the loop: just hooks, a
local evidence ledger, and exit codes.
- Why it exists — a USENIX Security '25 study found **5.2–21.7% of LLM-recommended packages dong" attack surface); grounded closes it at the moment of
install - How — a PostToolUse hook accrues evidence (files read, URLs fetched, registry checks) into a session ledger; a PreToolUse hook blocks actions that have none, feeding the reason back so the agent self-corrects
- Engineering — Python stdlib only, 210 offline tests, false-positives-are-worse-than-misses der path
A NeuralProphet + XGBoost hybrid that predicts the optimal purchase moment for airfare.
University AI project built on 18,472 rows of real fare data I collected daily via SerpApi — answering "buy now, or wait?"
Model NeuralProphet (per-route seasonality, base fare) + XGBoost (residuals: lead time, carrier, holidays final price = exp( log(base fare) + XGB_log_residual )
Performance (18,472 rows / 6 Korea–Japan routes)
| Evaluation | R² | MAE |
|---|---|---|
| TimeSeriesSplit (production) | 0.713 | ₩34,017 |
| GroupKFold NO-lookup (academic) | 0.673 | ₩34,879 |
- Daily automated collection via GitHub Actions → accumulating data pipeline
- Conformal Prediction (CQR) for 80% prediction intervals — 81.2% coverage achieved
- Streamlit dashboard visualizing BUY / WAIT / DROP action plans
- SHAP explanations per prediction (carrier, holiday, FX contributions)
Frontend developer at the startup Ohble.
- Built a React-based promotion website and a data-analytics dashboard
- Shipped and operated services in production
Tools & Infra
