# 윤수환 Suhwan Yoon

**AI Engineer** — Seoul, South Korea

> Five years of AI/NLP engineering on financial data (filings, news, stocks), taking LLM agents, RAG, and document extraction from research through deployed APIs. Currently building search and multi-agent systems at Wanted Lab.

Career Archetype: **Junction Artisan** (HE-MX) — One of the first Junction Artisans on Saywise

## Links

- LinkedIn: https://www.linkedin.com/in/%EC%88%98%ED%99%98-%EC%9C%A4-72110a168
- Career Mate: https://www.wanted.co.kr/ai/agent/career-mate

## Experience

### AI 기술팀 정규직, (주)원티드랩 (2025-11 – present)
Seoul, South Korea · Replaced the BM25 position search engine with a SPLADE + Dense + Lexical 3-leg hybrid search, tuned via A/B tests. Post-search bounce rate fell from 35% to 28% and search-driven application conversion rose from 6.7% to 8.4%. Also built the career mentor multi-agent service, where resume diagnosis raised user resume levels, and migrated legacy position agent code to LangChain and LangGraph.

### AI Engineer → AI Engineer Leader, 웨이커코리아 주식회사 (2022-10 – 2025-08)
Seoul, South Korea · Built a multi-agent AI chat API that uses MCP tools grouped per agent, LLM orchestration, and a GuardRail module blocking sensitive identifiers. Developed a financial stock analysis agent using Plan-and-Act workflows, a Llama-Index QA agent that validates extracted SEC filing data, and a Kafka-based news pipeline with translation and summarization APIs. Earlier work covered SEC filing extraction with LLM and multimodal models, plus model compression for translation (distillation, pruning, quantization, ONNX).

### NLP Developer 정규직, (주)디앤아이파비스 (2021-07 – 2022-10)
Seoul, South Korea · Built a BM25 patent document search engine on self-hosted Elasticsearch. Domain-tuned Sentence-BERT for problem and solution sentence matching, improving accuracy by 1.2% over vanilla BERT. Developed an LDA-based keyword extraction and clustering API for patent documents.

### 인턴 연구원, 네이버 주식회사 (2021-02 – 2021-05)
Seoul, South Korea · Researched training methods for ColBERT in information retrieval. Used Learning-to-Augment to generate hard negative samples and improved the Triplet Loss training approach.

## Education

### 컴퓨터공학과 석사 (자연언어처리), 경희대학교 대학원 (2019-03 – 2021-02)

### 컴퓨터공학과 학사, 경희대학교 (2013-03 – 2019-02)
2018 Junior Paper Competition Encouragement Award (KSC)

## Projects

### Career Mate 커리어 멘토 에이전트
Multi-agent career mentoring service at Wanted · Six workflows and a multi-agent system covering mentoring, mentor recommendation, resume diagnosis, and nudge message generation. Resume diagnosis raised data quality for 11,626 cumulative users.

### 번역 모델 경량화
Compressing MBART for service deployment · Built a half-depth MBART with knowledge distillation, then applied PTQ and attention-based pruning. Exported to ONNX as a static graph for serving.

### Question Difficulty Estimation (Applied Sciences)
Attention-based question difficulty model for QA · Co-first-authored a paper on estimating question difficulty with an attention model for question answering.

## Skills

LangChain, LangGraph, Llama-Index, Agentic AI, Multi-Agent, MCP, Prompt Engineering, GuardRail, Hybrid Search, SPLADE, BM25, OpenSearch, Elasticsearch, FAISS, Reranking, A/B Testing, PyTorch, Scikit-Learn, Named Entity Recognition, Topic Modeling, Knowledge Distillation, Model Pruning, Post Training Quantization, ONNX, Python, FastAPI, Flask, PostgreSQL, Redis, Apache Kafka, AWS, Docker, Jenkins, GitHub Actions, Git

## AI Fluency

- **LangChain** (https://langchain.com)
- **Claude Code** (https://claude.com)
- **Llama** (https://llama.com)
- **Hugging Face** (https://huggingface.co)

---
Source: https://saywise.com/member8228 (last modified 2026-10-27T05:59:14.580Z)