Backend engineer at 한국리서치. I ship LLM and agent systems into real workflows, and the part I care about is the part that comes after generation — the gate, the trace, the audit line that tells you whether the answer can be trusted.
tuto — evidence-tracked video understanding
Python · MIT · Claude Code plugin
The values that matter in a tutorial are often never spoken. They sit on screen, and
auto-captions get them wrong — 1536 becomes 136, and an agent that follows the
caption builds 136-dimensional vectors that fail on every call.
tuto reads the frames. Every knowledge item carries a timestamp, and items grounded in actual pixels are counted separately from items paraphrased off captions, so you can see which is which. When captions and screen disagree, the screen wins and the conflict is kept as structured data rather than smoothed into prose. Merging, cross-checking and the audit line are deterministic Python; exactly one call goes to the model.
I chose the architecture by measuring it rather than guessing. Six configurations on the same 54-minute video over one day: $11.12 → $4.75 (−57%), screen-grounded settings 12 → 40. Three results ran against my intuition, and measurement won each time. Benchmark run: 171 knowledge items, 69 screen-grounded, 0% uncited. 785 regression tests.
The limitations are in the README too, including the ones I have not solved.
Search-Pro — schema-grounded Text-to-SQL
C# · Python demo
Every data question in the team became a SQL request aimed at a developer. I led the internal AI contest team as its only engineer, defined the requirements with the people who actually run the queries, and shipped it.
The design rule was that the model never emits SQL. It returns a structured plan; the server renders the SQL; a read-only validation gate decides whether it runs. Failures are sorted into eight types and kept as regression tests. Quote work went from about 40 minutes per case to roughly 4, across ~490 cases a year.
star-coach-analysis — explainable swing coaching
Python · HCI Korea 2026
Pose-correction models hand back a fixed skeleton and no reason. STAR-Coach traces the optimization that produced the correction and generates the explanation from it: which body part, at what point in the swing, by how much. This repo is a mock-data reimplementation of the correction-analysis pipeline from the paper.
evaluation-awareness-transfer — research proposal
proposal stage — no experiments run yet
Recent work shows small LLMs internally distinguish "I am being evaluated" from ordinary use, readable with a linear probe — all of it in English. The proposal asks whether an English-trained probe survives Korean and Japanese input in 1–4B open models, and measures an in-language ceiling alongside it so that transfer failure can be told apart from the representation simply not being there.
Single researcher, ~9 hrs/week, one RTX 3060. The scope is set to fit that.
Explainable AI-powered Baseball Swing Coaching System Seunghyun Oh, Junsoo Hwang — Proceedings of HCI Korea 2026, pp. 1518–1523, Feb. 2026. Equal contribution.
한국리서치 — Backend Engineer (2024.09 – present), 온라인패널조사부
ASP.NET Core and Web Forms, SQL Server schema and query work, IIS on Windows Server. Legacy code, operational risk, unclear requirements, verification cost. That environment is where the habit came from: before shipping anything a model produced, ask what evidence would make this output checkable.
C# Python Java TypeScript · ASP.NET Core Spring Boot React · MSSQL IIS Nginx
Making agent output verifiable is the thread through all of the above, and it is what I want to keep working on — whether an explanation is faithful or merely plausible, which parts of an agent trace actually help someone debug a failure, and how to catch SQL and other generated artifacts that are wrong but entirely convincing.
Contact · jsjsjsjsjs1019@naver.com · Portfolio · Notion Portfolio