Projects

A Windows 11 and WSL2 terminal that brings split panes, per-pane tabs, workspaces, and coding-agent status into one interface.

Personal ProjectDeveloper Tool

Role

Solo · Product planning, architecture, implementation, release, and maintenance

Stack

Rust · Tauri v2 · TypeScript · xterm.js · WebView2 · ConPTY · WSL2 · GitHub Actions

Source code →

Compared a random baseline with three curriculum conditions across two learning-rate settings in eight training runs.

Independent Research

Role

Solo · Experimental design, data, training pipeline, and evaluation

Stack

Python · PyTorch · Hugging Face · kiwipiepy · RunPod

A personal site that manages projects and articles in MDX and deploys statically to GitHub Pages.

Personal Project

Role

Design · Development · Operations

Stack

Next.js 16 · React 19 · TypeScript · Tailwind CSS v4 · MDX · GitHub Actions

Source code →Live demo →

Building collaboration and cloud infrastructure, along with automation tools for architectural workflows, at an architecture firm of roughly 100 people.

Professional WorkIn Progress

Role

AA Architects & Planners · Head of AI Lab — first dedicated digital and AI transformation lead and in-house developer

Stack

Python · FastAPI · PostgreSQL · Docker · Terraform · GCP · MCP · LLM API

Used constraint optimization to restore the valid question pool from about 40,000 to about 90,000 after an answer-choice reuse bug.

Professional Work

Role

BookIPs · AI Engineer

Stack

Python · SQL · dbt · PostgreSQL · Prefect · LLM · Integer Programming

A PoC comparing a lower-cost commercial model with log-probability classification to improve consistency in item evaluation.

Professional WorkPoC

Role

BookIPs · AI Engineer

Stack

Python · LLM API · Log Probability · Prefect · Batch Evaluation

Systematized type-specific generation and validation criteria, raising the initial internal review pass rate from under 10% to about 70%.

Professional WorkProduction

Role

BookIPs · AI Engineer — Designed type-specific generation and validation logic and led iterative experiments

Stack

Python · FastAPI · PostgreSQL · Prefect · dbt · GCP · LLM API · ML Serving

A web tool that generates LangGraph code matching a graph laid out visually with nodes and edges.

Personal Project

Role

Solo · Product design, implementation, and deployment

Stack

React · React Flow · LangGraph · Netlify

Source code →Live demo →

Classified credit card customers from 857 features, placing 9th overall (top 4%) as a solo competitor.

CompetitionSolo Entry

Role

Solo · AI Hub × DACON competition

Stack

Python · XGBoost · Optuna · Weights & Biases

Built a model that translates Korean sentences into Korean Sign Language morphemes and diagnosed failures in the 3D avatar integration.

Personal Project

Role

3-person team · EDA, translation-model fine-tuning, and Blender-based error diagnosis

Stack

Python · koBART · PyTorch · WandB · Blender · Unity

Encoded a patent attorney's drafting criteria as LangGraph business logic and delivered a containerized module with explicit failure and retry paths.

Professional WorkContract

Role

Sole engineer responsible for the AI module (startup contract)

Stack

Python · LangChain · LangGraph · LLM · Docker

Condensed patent-dispute response reports averaging 200 pages into one or two pages within HyperCLOVA X's 8K-token limit.

Industry Collaboration

Role

4-person team · LLM summarization module and keyword extraction (industry collaboration)

Stack

Python · HyperCLOVA X · Embedding · MySQL

Extended a gain from query rewriting into a document-rewriting hypothesis, improving retrieval and placing second in the competition.

Competition

Role

2-person team · Query/document rewriting experiments and embedding retrieval (Upstage AI Lab competition)

Stack

Python · Embedding · TF-IDF · HNSW · LLM

Compared fine-tuned summarization models and LLMs to place 1st in a dialogue-summarization competition.

Competition

Role

3-person team · Data augmentation, summarization models, and LLM fine-tuning (Upstage AI Lab competition)

Stack

Python · OpenAI · Fine-Tuning · KoBART · QLoRA

Classified 17 document types and used OCR only for low-confidence predictions, placing 5th in the competition.

Competition

Role

4-person team · EDA, model training, and OCR (Upstage AI Lab competition)

Stack

Python · ConvNeXt · EfficientNet · ViT · Augraphy · OCR

Used architectural domain knowledge and external data to fill missing values, placing first in the competition.

Competition

Role

4-person team · EDA, feature engineering, and external data collection (Upstage AI Lab competition)

Stack

Python · XGBoost · QGIS · Iterative Imputer