AI/ML Engineer
I build AI systems that run in production: LLM and vector search, data pipelines on AWS, and MLOps.
I have spent 2+ years building and running AI/ML systems. Most recently I was the only AI/ML engineer at an online store for used auto parts. Before that, I led the AI team at a renewable-energy company.
I live in Daejeon, South Korea, and I am looking for AI/ML engineering roles in South Korea first. I plan to move to Germany later. I am also open to roles in Germany and elsewhere in Europe, including Switzerland and Denmark, and to remote roles.
Selected work
Four projects that show how I work. Each case study tells the problem, what I built, how I measured it, and what it changed. The pages leave out employer code and private data.
Buyers type a plain description in their own language, and the search finds the exact part for their exact car. I combined vector search, entity recognition, and an LLM that cleans up the query, and built an automated accuracy test suite.
It scored 95%+ on the 3,000-part test of a government R&D program, and the project was approved.
Pricing research needed a daily feed of used-parts listings from the market. I built an event-driven pipeline: on a schedule, Lambda workers take jobs from an SQS queue and save the listings to PostgreSQL, and CloudWatch raises an alarm when something fails.
It reached its target of 5M listings a day in testing, on the smallest database instance. In production, the source's rate limits held it to about 5,000 a day.
I led the team that put two models into production: one forecasts how much power a solar plant will make, and one flags unusual readings in the plant's data. Both fed a Virtual Power Plant, a platform that runs many plants together as one.
Forecast error fell from 9% to 7%, enough to sell power on the Korea Power Exchange. The detector's precision rose 35 points and its recall 37. A deploy took 15 minutes instead of 3 hours.
The data team loaded third-party catalog data by hand. I built an agent workflow that takes a brand, model, year, and region, fetches the matching parts data, cleans it, and writes it into the catalog. Jobs run on a schedule, resume after a failure, and end with a report the team can check.
A working prototype with a demo UI for the data team. It was not in production when I left.
I researched a model that prices a used part from depreciation, demand, and live market signals. On a small set of real prices it was about 70–80% accurate, measured with MAE and MAPE.
I built an LLM pipeline that wrote an English and a Korean summary for each of the company's ~40K products, following rules from the parts experts. It stayed a proof of concept.
I moved ~40K company records to a new database and loaded 2M+ third-party catalog records. Duplicates were removed, and each batch was written atomically, so a failure could not leave half-written data.
How I build
Since April 2026 I write most software by directing coding agents, Claude Code and Codex. The agents are fast. The rules, project notes, and checks I put around them keep the work reliable. Two apps I use myself and a physics engine anyone can play show the result.
The app reads my bank's Excel statement on my phone. Code reads every amount and date, AI picks each category, and I check each row. Then my girlfriend gets a Telegram message with what she owes.
I estimate it saves me 40 to 100 hours a year.
I describe a conversation in plain words, and an AI agent saves it on the right people and answers questions about them. It acts only through tools that run on my phone, and deletes wait for my tap.
One tap undoes a reply, and every wrong answer can be reported for fixing.
Skin, muscle, and bone respond to each blow through real physics, right in the browser. I can't judge physics numbers, so each test names the real outcome and the study behind it.
Every change replays each test swing 45 ways before it goes live.
I keep one central set of rules and project notes, and 19 projects load it: 9 at work and 10 personal. Claude Code and Codex read the same standards and pick up the context of earlier sessions. An agent that cannot load the rules must stop and ask.
The rules make agents keep changes small, update the docs, and never read secrets or touch production databases. A task counts as done only after a real end-to-end check passes.
A small service, backed by Claude, that gives all my apps one AI endpoint. It accepts OpenAI-style requests, so existing clients work unchanged, and it returns output that matches a JSON Schema.
It can also run an agent loop: the model decides which tool to call, but the calling app runs the tool on its own data and sends back only the result. The gateway never stores app data.
Besides Expenses and Network App, the agents built OneShot Reminder, Notes, and Life Simulator. OneShot Reminder also uses AI, and it asks me to confirm each task before it saves. The apps update themselves from signed GitHub Releases.
Earlier ML work includes image classifiers that grade diabetic retinopathy, built for freelance clients, and reinforcement-learning experiments for solar forecasting. More is on GitHub.
Experience
Mar 2025 – Jul 2026 South Korea (remote)
G-Man Auto Parts · used auto-parts e-commerce
Nov 2024 – Feb 2025 Remote
Upwork · medical image analysis
Mar 2024 – Nov 2024 Naju, South Korea
Recs Innovation · renewable energy
Promoted twice: Data Science Intern → AI Engineer → AI Team Lead
Skills
Education & credentials
The small images are scans of the diplomas and award certificates, with the birth date covered. Open one to see it full size.
Contact
I'm looking for AI/ML engineering roles: LLM and search systems, MLOps, AI backend, and data platforms. Email is the fastest way to reach me.