CV

Kwangmin Kim (김광민) · AI Engineer / Data Scientist · Seoul, South Korea kmink3225@gmail.com · GitHub · LinkedIn

Summary

AI Engineer / Data Scientist with 7+ years of experience, architecting and building enterprise AI platforms (RAG, LLM agents, NLP) end-to-end. I built an enterprise AI-agent knowledge platform from the architecture up (~98% user satisfaction) and delivered a data standardization system (validation time cut 99%) that is now expanding into a company-wide multi-agent platform under my technical lead. My self-built agent orchestration benchmarked up to ~18× lower cost than general-purpose models; I also cut diagnostic-equipment QC operating cost ~13×/yr, alongside statistically rigorous model evaluation and experiment design. I have led multidisciplinary teams (up to ~20) and filed 7 patents (first inventor on 4).

Specialties: LLM agents, RAG system design/implementation, NLP / deep learning, machine learning, experiment design, statistical analysis, diagnostic algorithms.

Experience

Seegene — Data Scientist / AI Engineer

Diagnosis IT General Research Institute · Data Science / Core Dev Team · 2020.12 – Present · South Korea

Enterprise AI-Agent Knowledge Platform — Technical Lead / AI Architect, 2025.11 – Present

• Led end-to-end architecture of a domain-specific multi-agent RAG platform for company-wide data assetization (deployed to working-level staff, expanding company-wide) — three agents (knowledge QnA, data standardization, sequence-recommendation code analysis) on shared Azure infrastructure; scaled a single-agent plan into a multi-agent flagship and delivered two agents 2 months ahead of target.
• Knowledge QnA chatbot — 9 sub-agent Self-RAG/CRAG loop with token streaming and source citation over a Parent-Child + hybrid-search (BM25 + vector) RAG pipeline; passed all 10 metrics (4.66s avg response, 96.9% citation rate, 100% system success, 95.6% retrieval success), 5.0/5.0 factuality & reasoning (gpt-4.1) on a 4-model LLM-as-judge eval, ~98% user satisfaction.
• Data-standardization assistant agent — Rule + ALBERT classifier + RAG hybrid engine (LangGraph Reflexion loop) auto-recommending three metadata types; passed all 10 metrics, 90.4% satisfaction, 3.75s avg response, 0% fallback.
• Sequence-recommendation code-analysis agent — grounded ~400K lines of Python (32 repos, 1,453 files) into 40K AST facts, a code graph (11,729 nodes / 38,783 edges), and a 42K search index; benchmarked three architectures (raw general-purpose CLI vs. metadata+skill harness vs. self-built orchestration; 7 variants) on a 6-metric composite + statistical tests — the harness beat the general CLI on answer usefulness (cross-validated by blind practitioner review), and the self-built orchestration won overall (GPT-5.4-mini composite 0.977, 11.6s, $0.076/query), ~17× cheaper than the costliest variant; nearing production.
• Evaluation & MLOps baseline — LLM-as-judge auto-scoring (factuality / reasoning / out-of-scope / multi-turn) + architecture A/B benchmarking (paired t-test, McNemar, Cohen’s d, bootstrap CI) + metric logging; ran 32% below projected cloud operating cost, with a self-built harness strategy hedging vendor lock-in.
• Drove two Microsoft workshops, persuading an MS architect and 7 engineers to adopt the self-built orchestration over a general-purpose Copilot CLI.

NLP-Based Data Standardization System — Technical Lead (mentored 20+ across IT/BT), 2024.10 – 2025.09

• Defined the metadata-inconsistency problem and led an NLP + Rule + RAG standardization system end-to-end; after a successful pilot it went company-wide and seeded the follow-on AI-agent platform.
• Outcomes (user survey + ops): validation time 8h → 0.73s (99%↓), cross-team inquiries 70 → 4/mo (94.3%↓), metadata consistency 8.4% → 98.7%, completeness 29.6% → 100%.
• 8-model classifier benchmark (KLUE-RoBERTa, XLM, KoBERT, ALBERT, mBERT, BiLSTM, DistilKoBERT, e5; 14 classes, 7,698 samples, stratified, 95% CI, McNemar+Holm over 28 pairs) → KLUE-RoBERTa 96.88% (top-5 transformers statistically tied).
• Robustness / 5-way cross-validation — 5-fold CV showed a 671K-param BiLSTM statistically on par with the 110M KLUE (96.18%±0.41% vs. 96.35%, p=0.73) at 1.48ms inference (vs. 12.49ms); suffix ablation (−51%p), a RAG holdout (rejected synthetic-overfit), and a noise floor diagnosed the accuracy ceiling as a data limit.
• Training-data engineering — curated 9,168 items from three sources (LLM, rules, RAG) → label normalization, 29 conflicts resolved, 1,466 deduplicated → 7,698; built dictionaries of 582 standard terms and 147 domain mappings.
• Rule-based naming-standardization engine (14 rules + physical-name / abbreviation generation); synonym clustering (ko-sroberta-multitask + HDBSCAN, 2,048 → 569 clusters); pytest (60), GitHub Actions CI, Docker for reproducible ML.

Time-Series PCR Signal Baseline-Correction Optimization — Project PM (DS 3, DE 1), 2024.01 – 2024.09

• Redesigned a hard-coded legacy baseline algorithm into a mixed-basis data-driven model, cutting the false-negative rate 0.47% → 0.04% (91.49%↓); refactored Matlab → low-level Python with real-time lightweight regression, ranking 1st of 5 competing algorithms on white-noise residual fit.

FDA-Submission Diagnostic-Algorithm Safety Statistical Analysis — Project PM (16, multidisciplinary), 2023.05 – 2023.12

• Designed and automated the statistical V&V pipeline for FDA software validation, cutting validation time 6 months → 3 weeks (87.5%↓) at 99.2% statistical confidence; implemented C++-port statistical tests (2-way RM-ANOVA, McNemar, Breslow-Day, Cochran-Mantel-Haenszel), an in-house Switch Model ablation, and an Airflow → R + Quarto pipeline generating a 200-page V&V report.

RT-PCR Diagnostic-Algorithm Reverse Engineering & Statistical Modeling — Data Scientist (team of 6), 2021.10 – 2023.04

• Reverse-engineered an undocumented legacy Matlab algorithm (10+ stages, 50+ empirical parameters) to 80% logic/dependency coverage with a C++-port spec; designed an RT-PCR-kinetics logistic-sigmoid composite with joint normal estimation to remove systematic bias.

PCR-Equipment QC Protocol Design & Performance Grading (A+/A/B/F) — Project PM (11, multidisciplinary), 2020.12 – 2021.09

• Automated manual Excel QC with a two-stage LSTM + 10 quality metrics grading system: QC time ~400h → 28h per 100 units (93%↓), ~13× annual operating-cost reduction; over 2,201 units and 61,248 signals, 94.5% pass/fail and 82.7% grade accuracy, with PCA/t-SNE/DBSCAN anomaly detection and an R Shiny dashboard — R&D President’s Award, 2 first-inventor patents.

Columbia University Irving Medical Center — Taub Institute · Statistical Research Assistant

Research on Alzheimer’s Disease and the Aging Brain · 2018.12 – 2020.05 · New York, US

• Integrated genomic, metabolomic, and clinical data to surface 13 key biomarkers (p<0.01) from ~3,000 metabolites, resolving a confounder missed for eight months.
• Worked a high-dimensional, small-sample regime (146 samples × 3,000 variables); compared 10+ ML algorithms and chose sPLS (84% accuracy with interpretability); built 20-year onset-risk models with Cox hazard and family-based GEE — research-competition top 3, Chair’s Award, full-time neurosurgery offer.

Education

• M.S. Biostatistics, Columbia University (2017–2019) — Chair’s Award (annual graduation research competition)
• B.A. Mathematics, Baruch College, CUNY (2015–2017)
• B.S. Biochemistry, Kangwon National University (2006–2012) — Valedictorian, Dean’s Award

Skills

• LLM Agent / GenAI — RAG, Agentic RAG, Graph RAG, Self-RAG/CRAG, LangChain, LangGraph, Azure OpenAI, Azure AI Search, OpenAI/Claude API, Prompt Engineering
• NLP / Deep Learning — KLUE-RoBERTa, KoBERT, ALBERT, KoSRoBERTa, BiLSTM, LSTM, Hugging Face Transformers, PyTorch, KiwiPiePy/KoNLPy
• ML / Statistics — scikit-learn, HDBSCAN, regression/survival analysis, time series, causal inference (A/B test), experiment design, 95% CI, McNemar/Holm
• Data / Backend Engineering — Python, R, SQL (MySQL/PostgreSQL), SAS, FastAPI, Streamlit, Apache Airflow, Parquet, AST
• Visualization / Docs — Plotly, Matplotlib, Seaborn, R Shiny, ggplot2, Quarto, Jupyter, R Markdown
• Cloud / DevOps — Azure, Azure DevOps, Docker, Git/GitHub, Conda

Patents (filed)

• (First inventor) Customized treatment method based on repeatedly-measured Ct values, Seegene (2022)
• (First inventor) Subscription system for a medical platform, Seegene (2022)
• (First inventor) Noise-test automation system for diagnostic equipment, Seegene (2021)
• (First inventor) Noise-level measurement algorithm for medical equipment, Seegene (2021)
• (Co-inventor) Prediction model for molecular diagnostics, Seegene (2022)
• (Co-inventor) Negative certificate for molecular diagnostics, Seegene (2022)
• (Co-inventor) Molecular-diagnostics system for community groups, Seegene (2022)

Awards & Certifications

• President’s Award (R&D) — noise-test automation system, Seegene (2021)
• Chair’s Award — Graduation Practicum Research Competition, Columbia Biostatistics (2019)
• Job Offer — Taub Institute, Columbia University Irving Medical Center (2019)
• Dean’s Award — valedictorian, Kangwon National University (2012)
• Microsoft Azure certification training — DP-203 (Data Engineering), DP-100 (Data Science), DP-300 (Database) (2025)
• SAS Certified Base Programmer (2018), SIT TESOL Instruction Certification (2014)
• Completion — EN62304 Medical Device SW Life Cycle (SGS, 2021), HIPAA (CUIMC, 2020)
• Stipends — $1,000 Mathematical Kinetic Modeling, CUNY (2015); $5,000 Medical Convergence Capstone Design, KNU (2012); full academic-excellence scholarship, KNU (2010–2011)

Teaching & Mentoring

• Mentor, Seegene — AI Engineering (2024–2025), Data Standardization (2025), Statistical Analysis (2023–2024), Intro to Statistical Learning (2022)
• Teaching Assistant, Columbia University — Probability Theory (graduate, 2019)
• Teaching Assistant, CUNY — Calculus 1–3, Precalculus, Statistics (undergraduate, 2015–2016)
• Private Tutor — Calculus 1–2 (New York, 2021), GRE Math, TOEFL iBT (New York, 2014–2020)
• Trainee Instructor — SIT TESOL teaching, Rennert (2014)

Languages

• Korean (native) · English (fluent)