AI Product Engineer · Seoul

I make AI
work for real.

Software engineer with a data engineering foundation, building and shipping production AI agents across enterprise, customer-support, and research contexts.

See selected work LinkedIn

Previously
Sendbird · BCG · LG AI Research

+21%

CSAT lift
Production support agent

108×

Faster evaluation
9 hours → 5 minutes

−70%

Cloud storage cost
Distributed data pipeline

01 / Selected work

Systems that moved beyond the demo.

I work across product, data, and AI boundaries—especially where reliability, messy operational context, and real users make the problem interesting.

Sendbird · 2025—2026 01

Production CS AI agent

Improved a live AI agent by tracing failures through the full system.

Worked on a customer-support agent for a major food-delivery platform, from tool routing and proxy architecture to batch evaluation.

  • Improved CSAT from 2.9 to 3.5 during a 5% regional rollout.
  • Reduced evaluation review time from 9 hours to 5 minutes across 1,000+ conversations.
  • Built an internal Slack, Jira, and GitHub issue tracker adopted by the team.
AI Agents Tool Routing LLM Evaluation Production Debugging
Boston Consulting Group · 2025 02

Enterprise AI · Multi-agent RAG

Traced failed queries to the data-routing layer.

Built a ReAct-based multi-agent RAG system for enterprise sales data, then found RDB-to-VectorDB misrouting and data-integrity issues behind failed queries.

LangGraph RAG PostgreSQL System Debugging
LG AI Research · 2023—2024 03

AI data infrastructure

Rebuilt the data layer underneath model development.

Scaled Korean web data collection with Kubernetes and Redis, cut cloud storage cost by 70%, and built retrieval datasets for LLM tool planning.

Kubernetes Redis Airflow Vector Search

02 / How I work

From spoken context to traceable execution.

My real workflow starts in meetings. I turn raw conversation into durable context, then use that context to create concrete work—without manually copying it across tools.

  1. 01 / Capture

    Alt

    An STT-powered desktop app captures meeting transcripts so decisions and technical context do not disappear when the call ends.

  2. 02 / Structure

    Claude Code Skill

    A custom skill turns each transcript into a structured meeting note: decisions, open questions, owners, and the context behind them.

  3. 03 / Accumulate

    Obsidian

    Meeting notes become linked, searchable project memory—building context over time instead of resetting with every conversation.

  4. 04 / Operationalize

    Skills + Linear MCP

    Another skill reads the meeting notes and creates scoped Linear tickets with the relevant decisions and context already attached.

  5. 05 / Execute

    Build + Verify

    Claude Code and Codex help execute the scoped work; tests, traces, and evals close the loop before anything ships.

Working principle

The goal is not “more agents.” It is less context loss—from what the team said, to what we decided, to what actually ships.

03 / Experience

Data engineering roots. Production AI trajectory.

Recognition

Minister of Science & ICT Award · The Big Contest 2023

1st place / 740+ teams · Dec 2023

Gold Prize · The 5th Open Infrastructure Development Competition

Aug 2023

Grand Prize · The Regional Public Safety Data Analysis Contest

Feb 2023

04 / Toolkit

Enough range to follow the problem end to end.

AI systems

  • AI Agents
  • RAG
  • LLM Evals
  • LangGraph
  • Tool Use
  • Human Review

Data + backend

  • Python
  • SQL
  • FastAPI
  • Airflow
  • Redis
  • PostgreSQL
  • BigQuery

Infrastructure

  • Docker
  • Kubernetes
  • GCP
  • AWS
  • CI/CD
  • Observability

Daily workflow

  • Alt
  • Obsidian
  • Linear
  • Claude Code
  • Codex
  • Skills
  • MCP
  • GitHub