Sangryul Kim

I’m Sangryul Kim, a Machine Learning Engineer at NAVER. I design and build Agentic AI systems, and right now that means LLM evaluation systems and agent prototypes for the content feeds in the NAVER app. Someone tells me the result they want; I work out an architecture that gets there and keep reshaping a working version from their feedback. Many of the people I build for don’t write code, so I start from the outcome they care about and check whether the system reaches it.

How much room the model gets decides the architecture. Where the rules are clear, a single judge with a strict rubric does the job. Where the model has real choices to make, a planner proposes, code holds it to a fixed schema, and a blind judge scores the result without knowing which strategy produced it; when one viewpoint isn’t enough, I split the work across cooperating agents, such as persona agents that each read an A/B test for one user segment while a synthesizer combines them. I write that orchestration in plain Python rather than starting from a framework.

Retrieval is where I started. I researched RAG and question answering at KAIST, where I received my M.S. in AI in Feb. 2025 with Prof. James Thorne in the XFACT lab. In my last months there I interned at NAVER Search, training a Korean text embedding model, and afterwards designed a pipeline for LINER’s agent search service in which an LLM agent decides what to search for and gathers what it needs to answer a question. In the NAVER app I still build retrievers and run ranking experiments for the feeds.

Before graduate school, I interned as an ML engineer at NAVER WEBTOON and built the web service for the pre-release version of IMQA at ONYCOM. During my military service, I wrote software at the Cyber Operations Command. I studied Computer Science and Engineering at Chung-Ang University, with a minor in Applied Statistics.