content: sync CV + flagship + standardization pages with 2026-07-03 career-os master

- benchmark: 7 -> 11 variants, 3 architectures named (raw Claude Code vs
  metadata/skill harness vs self-built orchestration), 51-question eval set,
  $0.076 vs $1.32 (~17x)
- CV summary: ~18x -> ~17x, team size to SSOT (16 led / 20+ trained)
- standardization: period 2024.06-, Data Architect phase, commercial
  governance solutions (Databricks/Snowflake/Purview) build-vs-buy framing
- drift fixes: teaching years, HIPAA 2018, CUIMC intern->SRA

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: kmink3225

_pages/cv.md

@@ -16,7 +16,7 @@ toc:
 
 ## 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)**.
+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 ~17× lower cost** than general-purpose agents; 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 16), trained 20+ practitioners in a data-standardization framework, 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.
 
@@ -31,13 +31,13 @@ AI Engineer / Data Scientist with 7+ years of experience, architecting and build
 - 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.
+- **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 Claude Code vs. Claude Code + metadata/skill harness vs. self-built orchestration; **11 variants**, including the latest Claude Sonnet 5 both raw and harnessed) on a 6-metric composite + statistical tests over a 51-question eval set — the metadata/skill harness beat raw Claude Code on answer usefulness (cross-validated by blind practitioner review), and the **self-built orchestration won overall** (GPT-5.4-mini composite **0.977**, 1st of 11, 11.6s, $0.076/query), **~17× cheaper** than the costliest variant ($0.076 vs. $1.32); 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
+**NLP-Based Data Standardization System** — Technical Lead (Data Architect for the framework-design phase, 2024.06–09; mentored 20+ across IT/BT), 2024.06 – 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.
+- Defined the metadata-inconsistency problem; evaluated commercial governance solutions (Databricks, Snowflake, MS Purview) and found they all **presuppose an existing standardization framework**, so designed the framework itself first, then 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.
@@ -60,7 +60,7 @@ AI Engineer / Data Scientist with 7+ years of experience, architecting and build
 
 - 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
+### Columbia University Irving Medical Center — Taub Institute · Research Intern → Statistical Research Assistant
 
 *Research on Alzheimer's Disease and the Aging Brain · 2018.12 – 2020.05 · New York, US*
 
@@ -100,12 +100,12 @@ AI Engineer / Data Scientist with 7+ years of experience, architecting and build
 - **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)
+- **Completion** — EN62304 Medical Device SW Life Cycle (SGS, 2021), HIPAA (CUIMC, 2018)
 - **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)
+- **Mentor**, Seegene — AI Engineering (2025–2026), Data Standardization (2025), Statistical Analysis (2023–2024), Intro to Statistical Learning (2023–2024)
 - **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)

_pages/ko_cv.md

@@ -15,7 +15,7 @@ toc:
 
 ## 요약
 
-7년 이상 경력의 AI Engineer / Data Scientist. RAG, LLM Agent, NLP 기반 엔터프라이즈 AI 플랫폼을 아키텍처부터 설계·구축한다. 사내 문서·지식·규정 기반 지식 QnA 챗봇을 포함한 엔터프라이즈 **AI Agent 지식 플랫폼**을 아키텍처부터 구축(사용자 만족도 **~98%**)했고, **데이터 표준화 시스템**을 성공시켜(검증 시간 **99% 단축**) 전사 멀티 Agent 플랫폼으로 확장, 총괄 설계 중이다. 자체 Agent orchestration으로 범용 대비 비용 **최대 ~18배 절감**을 벤치마크로 입증했으며, 진단 장비 QC 자동화로 연간 운영비 **약 13배 절감**, 통계적으로 엄밀한 모델 평가·실험설계를 병행한다. 다학제 팀(최대 20여 명)을 리딩했고 특허 **7건(제1발명가 4건)**을 출원했다.
+7년 이상 경력의 AI Engineer / Data Scientist. RAG, LLM Agent, NLP 기반 엔터프라이즈 AI 플랫폼을 아키텍처부터 설계·구축한다. 사내 문서·지식·규정 기반 지식 QnA 챗봇을 포함한 엔터프라이즈 **AI Agent 지식 플랫폼**을 아키텍처부터 구축(사용자 만족도 **~98%**)했고, **데이터 표준화 시스템**을 성공시켜(검증 시간 **99% 단축**) 전사 멀티 Agent 플랫폼으로 확장, 총괄 설계 중이다. 자체 Agent orchestration으로 범용 대비 비용 **최대 ~17배 절감**을 벤치마크로 입증했으며, 진단 장비 QC 자동화로 연간 운영비 **약 13배 절감**, 통계적으로 엄밀한 모델 평가·실험설계를 병행한다. 다학제 팀 리딩(최대 16명), 표준화 체계 20여 명 교육, 특허 **7건(제1발명가 4건)**을 출원했다.
 
 **전문분야:** LLM Agent, RAG 시스템 설계·구현, NLP/딥러닝, 머신러닝, 실험설계, 통계분석, 진단 알고리즘.
 
@@ -30,13 +30,13 @@ toc:
 - **전사 데이터 자산화**를 위한 도메인 특화 **멀티 Agent RAG 플랫폼**을 아키텍처부터 총괄 설계·구축(현업 실무진 배포, 전사 확장 중) — 지식 파편화·암묵지 소멸 문제 해결, 3대 Agent(지식 QnA, 데이터 표준화, 서열 추천 코드 분석)와 공유 Azure 인프라; 단일 Agent 계획을 멀티 Agent 대형 과제로 확장·승격, 2종 Agent를 목표 대비 **2개월 조기 완료**.
 - **지식 QnA 챗봇** — 9개 sub-agent **Self-RAG/CRAG** 루프 + 토큰 스트리밍 + 출처 인용(Parent-Child + Hybrid Search(BM25+Vector) RAG 파이프라인); 10개 지표 전수 통과(평균 응답 **4.66초**, 인용률 96.9%, 시스템 성공률 100%, 검색 성공률 95.6%), 4모델 LLM-as-judge 평가에서 gpt-4.1 사실·추론 **5.0/5.0**, 사용자 만족도 **~98%**.
 - **데이터 표준화 도우미 Agent** — Rule + ALBERT 분류기 + RAG **하이브리드 엔진**(LangGraph Reflexion 루프)으로 메타데이터 3종 자동 추천; 10개 지표 전수 통과, 만족도 **90.4%**, 평균 응답 3.75초, Fallback 0%.
-- **서열 추천 코드 분석 Agent** — 약 40만 줄 Python(32개 레포, 1,453 파일)을 **40K AST 사실**, 코드 그래프(**11,729 노드 / 38,783 엣지**), 42K 검색 인덱스로 그라운딩; raw 범용 CLI vs 메타데이터+스킬 하네스 vs 자체 orchestration 3-아키텍처(7개 변형)를 6-지표 Composite + 통계검정으로 비교 — 하네스가 범용 CLI 대비 응답 유용성 향상(실무자 블라인드 평가로 교차검증), **자체 orchestration이 최우수**(GPT-5.4-mini Composite **0.977** 종합 1위, 11.6초, $0.076/건), 최고 비용 변형 대비 **~17배 저렴**, production 마무리 단계.
+- **서열 추천 코드 분석 Agent** — 약 40만 줄 Python(32개 레포, 1,453 파일)을 **40K AST 사실**, 코드 그래프(**11,729 노드 / 38,783 엣지**), 42K 검색 인덱스로 그라운딩; raw Claude Code vs Claude Code+메타데이터/스킬 하네스 vs 자체 orchestration 3-아키텍처(**11개 변형**, 최신 Claude Sonnet 5를 raw·메타데이터 하네스 양쪽 포함)를 6-지표 Composite + 통계검정(51문항 평가 셋)으로 비교 — 메타데이터/스킬 하네스가 raw Claude Code 대비 응답 유용성 향상(실무자 블라인드 평가로 교차검증), **자체 orchestration이 최우수**(GPT-5.4-mini Composite **0.977** 종합 1위/11, 11.6초, $0.076/건), 최고 비용 변형 대비 **~17배 저렴**($0.076 vs $1.32), production 마무리 단계.
 - **모델 평가·MLOps 체계** 구축 — LLM-as-judge 자동 채점(사실/추론/범위외/멀티턴) + 아키텍처 A/B 벤치마크(paired t-test, McNemar, Cohen's d, bootstrap CI) + 메트릭 로깅; 클라우드 운영비 추정 대비 **32% 절감**, 자체 하네스 전략으로 벤더 Lock-in 대비.
 - **Microsoft 워크숍 2회** 기술 주도 — Azure 기반 Copilot CLI 대비 자체 오케스트레이션 토론에서 **MS architect와 엔지니어 7명을 자체 아키텍처로 설득**.
 
-**자연어처리 기반 데이터 표준화 시스템 구축** — Technical Lead (IT/BT 20여 명 멘토링), 2024.10 – 2025.09
+**자연어처리 기반 데이터 표준화 시스템 구축** — Technical Lead (표준화 체계 설계 단계 2024.06–09엔 Data Architect; IT/BT 20여 명 멘토링), 2024.06 – 2025.09
 
-- 현업의 메타데이터 불일치 이슈를 **문제 정의** → NLP + Rule + RAG 기반 표준화 시스템 구축 총괄; 파일럿 성공 후 **전사 적용** 전환, **후속 AI Agent 플랫폼으로 확장된** 대형 과제의 출발점.
+- 현업의 메타데이터 불일치 이슈를 **문제 정의**; 상용 거버넌스 솔루션(Databricks, Snowflake, MS Purview)을 다각 검토했으나 모두 **표준화 체계가 이미 구축됐다는 전제의 적용 도구**임을 확인 → 표준화 체계 자체를 먼저 수립하고 NLP + Rule + RAG 기반 표준화 시스템 구축 총괄; 파일럿 성공 후 **전사 적용** 전환, **후속 AI Agent 플랫폼으로 확장된** 대형 과제의 출발점.
 - 자동화 성과(사용자 설문·운영 측정): 검증 시간 **8시간 → 0.73초(99%↓)**, 부서 간 문의 **월 70 → 4건(94.3%↓)**, 메타데이터 일관성 **8.4% → 98.7%**, 완전성 **29.6% → 100%**.
 - 도메인 분류기 **8-모델 벤치마크**(KLUE-RoBERTa, XLM, KoBERT, ALBERT, mBERT, BiLSTM, DistilKoBERT, e5; 동일 조건 14클래스, 7,698건, stratified, 95% CI, McNemar(Yates)+Holm 28쌍) → **KLUE-RoBERTa 96.88%** 최고(상위 5개 트랜스포머 통계적 동률).
 - **강건성·교차검증 5종** — 5-fold CV로 **671K BiLSTM이 110M KLUE와 통계적 동급**(96.18%±0.41% vs 96.35%, p=0.73) + **1.48ms 최속 추론**(12.49ms 대비) 발굴; suffix ablation(−51%p), RAG holdout(합성 과적합 가설 기각), 노이즈플로어로 정확도 상한이 데이터 한계임을 진단.
@@ -59,7 +59,7 @@ toc:
 
 - 수동 엑셀 QC를 **LSTM 2단계 예측 + 10개 성능 지표 등급 분류**로 자동화, QC 시간 약 **400h → 28h/100대(93%↓)**, 연간 운영비 **약 13배 절감(~8억원)**; 2,201대 장비·**61,248개 신호**로 합/불 94.5%, 등급 82.7% 분류, PCA·t-SNE·DBSCAN 이상탐지 + R Shiny 대시보드 — **R&D 부문 우수상(President's Award)**, **제1발명가 특허 2건**.
 
-### Columbia University Irving Medical Center (CUIMC) — Taub Institute · Statistical Research Assistant
+### Columbia University Irving Medical Center (CUIMC) — Taub Institute · Research Intern → Statistical Research Assistant
 
 *알츠하이머병·뇌 노화 연구소 · 2018.12 – 2020.05 · 미국 뉴욕*
 
@@ -99,12 +99,12 @@ toc:
 - **학장상** — 성적 우수 수석 졸업, 강원대학교 (2012)
 - **Microsoft Azure 자격 교육 이수** — DP-203(Data Engineering), DP-100(Data Science), DP-300(Database) (2025)
 - **SAS Certified Base Programmer** (2018), **SIT TESOL Instruction Certification** (2014)
-- **수료** — EN62304 Medical Device SW Life Cycle (SGS, 2021), HIPAA (CUIMC, 2020)
+- **수료** — EN62304 Medical Device SW Life Cycle (SGS, 2021), HIPAA (CUIMC, 2018)
 - **장학·Stipend** — $1,000 Mathematical Kinetic Modeling, CUNY (2015); $5,000 Medical Convergence Capstone Design, KNU (2012); 성적 우수 전액 장학금, 강원대학교 (2010–2011)
 
 ## 강의 · 멘토링
 
-- **Mentor**, Seegene — AI Engineering (2024–2025), Data Standardization (2025), Statistical Analysis (2023–2024), An Introduction to Statistical Learning (2022)
+- **Mentor**, Seegene — AI Engineering (2025–2026), Data Standardization (2025), Statistical Analysis (2023–2024), An Introduction to Statistical Learning (2023–2024)
 - **Teaching Assistant**, Columbia University — Probability Theory (석사 수준, 2019)
 - **Teaching Assistant**, CUNY — Calculus 1–3, Precalculus, Statistics (학부, 2015–2016)
 - **Private Tutor** — Calculus 1–2 (New York, 2021), GRE Math, TOEFL iBT (New York, 2014–2020)