Projects

Automated Korean and US stock trading where LLM agents in distinct roles split the work from research to order placement

KIS and Kiwoom automated trading built on a four-stage SimplifiedRunner with LangGraph Planning and Session Graphs on top. Zero Trust three-network isolation and staged BACKTEST→PAPER→CANARY→LIVE rollout.

A personal knowledge base RAG that Claude searches and indexes directly through MCP tools

A personal RAG that exposes Elasticsearch hybrid search (BM25+kNN+RRF), BGE reranking, and a concept graph to Claude through 14 MCP tools.

Markdown study material → LLM structuring → quizzes, spaced repetition, and annotation in one tool

A Next.js tool that structures markdown material into study content using a five-layer clean architecture, a Zod-validated LLM gateway, and the SM-2 (SuperMemo-2) spaced repetition algorithm.

Social matching built on mutual phone-number registration and a token economy

A social matching service that protects phone numbers with dual encryption and hashing and creates a Match through O(1) mutual-registration detection. Token economy, three payment providers, three social login providers.

The central Django backend for on-premises AI sorting systems at 20+ sites nationwide

An on-premises backend handling 30+ sorting requests per second with 8 Django apps, 5 Celery queues across 3 worker pools, and PostgreSQL RANGE partitioning.

Modbus TCP PLC communication middleware with four separated layers and a dedicated Celery polling worker

Four separated layers (Celery / Service / Modbus / Repository) and a dedicated polling worker collect data from up to 64 sorting outlets every 10 seconds.

A real-time GPU inference pipeline over 8 cameras plus a per-produce machine vision library

A torch.multiprocessing capture-inference-MV-transfer pipeline, paired with a standalone vision library package built on a BaseProcessor abstraction.

Remote monitoring of 20+ field sites, automated anomaly data collection, standardized CI/CD and deployment, GitHub-Slack-Notion webhook integration

Four-part operations infrastructure: Prometheus + Grafana monitoring, FastAPI anomaly collection, GitHub Actions + Docker Compose deployment, and a GitHub-Slack-Notion integration webhook.

A React + Electron desktop app running in kiosk mode on site

A React + Electron kiosk app holding 60fps on industrial PCs through virtual scrolling, adaptive polling, and monolith decomposition.

A four-phase port moving Flask Blueprints to Django + DRF and splitting the database

A four-phase port from Flask Blueprints to Django + DRF, splitting MySQL into PostgreSQL + MongoDB, alongside security fixes and a standardized three-layer architecture.

Leading the rollout of Git Flow, architecture docs, pre-commit, and Storybook across a five-person team

Established Git Flow with PR review, an architecture document set, a pre-commit quality system, and a Storybook component development environment across a five-person team.

An early prototype of an image-AI produce quality prediction system

An early produce AI sorting prototype built with Python, OpenCV, and PyTorch. It became the foundation for the systems built after the AIOFARM spin-off.