Projects
Company Projects
Samsung Electronics 사내 프로젝트 · 기밀 보호를 위해 세부 수치 일부 생략.
DRAM ML 최적화 · 양산 수율 혁신
ML Optimization Engineer · 2024–Present
🏆 대표이사상 (2025) — 삼성전자 전사 100,000명+ 중 단 3개 팀 선정. 역대 최고 수율 ML 자동화 성과.
- Server · HBM · Mobile 전 DRAM 제품 라인 ML 최적화 프레임워크 배포 → 역대 최고 수율 달성
- 차세대 DRAM 양산 프로그램 20개 이상 셋업
- 엣지 컴퓨팅 기반 ML 자동화 플랫폼 구축 (멀티디바이스 실시간 최적화)
- Bit Line Sense Amplifier(BLSA) 최적화 공정 구현 — Python 자동화 + C/C++ 서버 로직, 앙상블 알고리즘 통합
- 자동화 대시보드 · 모듈형 분석 툴 구축 → 생산 의사결정 간소화
- Chip-level MMT 데이터 기반 센싱 파라미터 자동 최적화 프레임워크 설계 (특허 출원 2건)
- Advanced Optimization Algorithms High-Grade Proposal 2건 제출 (2024)
특허 출원중 × 2
PythonC/C++ML OptimizationEnsembleEdge ComputingDRAMBLSAMMTDashboard
TB급 DRAM 데이터 플랫폼 구축
Data Analytics Engineer · 2022–2023
🏆 DREAM Fair Top10 (2023) · Development Best Performer (2023) — 1,000명+ 신입 엔지니어 중 자동화 툴 전사 배포 성과로 수상.
- 재사용 가능한 전처리 프레임워크 7종 개발 → DRAM 데이터 처리 시간 90% 단축, 전사 배포
- Dask · PyArrow · AWS Glue 활용 TB급 장애 로그 분산 빅데이터 파이프라인 아키텍처
- Airflow + Glue 트리거 기반 엔드투엔드 이상탐지 시스템 배포 및 리포팅 워크플로우 구축
- 공정/소자 전문가와 협업 → 분석 결과를 근본 원인 조사와 연결, 수율 개선 사이클 가속화
PythonDaskPyArrowAWS GlueApache AirflowAnomaly DetectionBigData PipelineSQL
반도체 공정 인과추론 · 하이브리드 ML 연구
Researcher · 2022–Present
🏆 Best Seminar Award (2023) — 'ML for Semiconductor Analysis' 강의 (2022–24, 사내 엔지니어 대상) 기여로 수상.
- Double Machine Learning(DML) 기반 Selective Interaction-Aware Causal Inference → 반도체 제조 공정 최적화
- DRAM 공정 파라미터 자동 튜닝을 위한 Hybrid ML 알고리즘 설계
- 사내 기술 저널 논문 3편 제출:
- Selective Interaction-Aware Causal Inference via DML for Semiconductor Process Optimization
- Hybrid Machine Learning Algorithms for Automated Semiconductor Process Optimization
- Optimizing DRAM Chip Treatment Selection by Enhanced Robust RL
Causal InferenceDouble ML (DML)EconMLHybrid MLPythonSemiconductorDRAM
Robust RL · DRAM 처리 전략 최적화
Researcher · 2022–2025 · IEEE IEEM 2025 (Melbourne)
🏆 Best Paper Award (2024) — Samsung Electronics Memory Business 사내 기술 컨퍼런스. "Optimizing DRAM Chip Selection via Robust RL" 연구.
- 순차적 의사결정 데이터 기반 DRAM 칩 처리 전략 최적화를 위한 Robust Reinforcement Learning 연구
- Decision Regime 개념 도입 → Progressive Decision-making 환경에서 모델 신뢰성 · 강건성 향상
- IEEE IEEM 2025 (Melbourne, Dec 2025) 구두 발표 — Accepted
- Samsung Product Engineering Technology Conference 2024 (Hwasung) 발표
- AISTATS 2026 후속 논문 투고 중 — Cross-Attention Temporal Fusion for Robust and Trustworthy Multimodal Learning
Reinforcement LearningRobust RLDecision RegimePythonPyTorchDRAMIEEE IEEMAISTATS
Program Projects
🧬 Alpha-Helix
AI 퀀트 투자 개인매니저 · 퀀트 개발자 전문 IDE · Claude Code
1
Backtesting Engine · LEAN + vectorbt
LEAN CI 프레임워크 + vectorbt 병행 · 7 전략 (SMA / RSI / MACD / Momentum / VIX / 무한매수법 / buy-and-hold) · QuantStats tearsheet 자동생성
LEAN CI 프레임워크 + vectorbt 병행 · 7 전략 (SMA / RSI / MACD / Momentum / VIX / 무한매수법 / buy-and-hold) · QuantStats tearsheet 자동생성
2
AI Signal Engine · Trust Score
XGBoost 13피처 + SHAP 설명 · 5-State HMM Regime · Trust Score 파라미터 섭동 · 22:30 자동 재학습
XGBoost 13피처 + SHAP 설명 · 5-State HMM Regime · Trust Score 파라미터 섭동 · 22:30 자동 재학습
3
KIS + Binance 실주문 자동화
MOCK → REAL 게이트 · OrderProposal 큐 · HMAC 서명 승인링크 · AES-GCM 자격증명 암호화 · 글로벌 Kill-Switch · KIS(한국주식) + Binance(암호화폐) 모두 실주문
MOCK → REAL 게이트 · OrderProposal 큐 · HMAC 서명 승인링크 · AES-GCM 자격증명 암호화 · 글로벌 Kill-Switch · KIS(한국주식) + Binance(암호화폐) 모두 실주문
4
Multi-LLM Daily Briefing (Gemini · Claude)
Gemini 2.5 + Claude 멀티 LLM 포트폴리오 코멘트 · Toss Payments FREE/STANDARD/PREMIUM 구독 · 3-process EC2 배포
Gemini 2.5 + Claude 멀티 LLM 포트폴리오 코멘트 · Toss Payments FREE/STANDARD/PREMIUM 구독 · 3-process EC2 배포
PythonFastAPILEANvectorbtXGBoostSHAPSpring Boot 4Java 21React 18Gemini 2.5ClaudeKIS APIBinance APIToss PaymentsMySQL 8AWS EC2
상세 내용 보기 →
🧬Alpha-Helix - AI 퀀트 투자 개인매니저 · 퀀트 개발자 전문 IDE
개인 프로젝트 · 배포 사이트: quant-alphahelix.com · 3-process 아키텍처 (Spring Boot + FastAPI analytics sidecar + React)
자연어 목표 → 전략 구성 → LEAN/vectorbt 백테스트 → QuantStats tearsheet → XGBoost+SHAP AI 시그널 → OrderProposal 큐 → MOCK→REAL KIS·Binance 실주문 자동화. Gemini 2.5 + Claude 멀티 LLM 일일 브리핑과 구독 플랜을 갖춘 실제 퀀트 펀드매니저 워크스페이스.
Backtesting Engine · LEAN CI + vectorbt · 7 Strategies
- LEAN CI 프레임워크. QuantConnect LEAN을 CI 백테스팅 엔진으로 통합 — 실제 퀀트 펀드매니저가 사용하는 리서치 파이프라인과 동일한 환경.
- vectorbt 병행 엔진. 7 전략 — buy-and-hold, SMA cross, RSI mean-reversion, MACD, momentum 12-1, VIX risk-off, 무한매수법. 0.25% commission + 0.1% slippage 반영.
- QuantStats tearsheet. HTML 리포트 자동생성 (
/reports/{file}.html) — Sharpe, Calmar, max drawdown, 월별 수익 히트맵 등. - Walk-Forward validation. 오버피팅 방지를 위한 롤링 윈도우 out-of-sample 검증 후 시그널 단계로 진입.
AI Signal Engine · XGBoost + SHAP + 5-State HMM Regime + Trust Score
- XGBoost up-probability. 13-feature model outputs daily up/down probability per ticker. Daily auto-retrain scheduler fires at 22:30 KST; Resilience4j Circuit Breaker + Retry wraps the analytics sidecar call.
- SHAP explainability. Per-prediction feature importance surfaces which indicators drove the signal — shown directly in the workspace Config/Report tabs.
- 5-State HMM Regime. Hidden Markov Model classifies market into 5 latent regimes (strong trend / weak trend / range / high-vol / crash) and gates strategy selection per regime.
- Trust Score & parameter perturbation. 4-factor reliability score stress-tests signal stability by perturbing model parameters — low-trust signals are flagged before reaching the order queue.
KIS + Binance 실주문 자동화 · MOCK → REAL Gate
- MOCK-first 아키텍처. 시그널 생성기는 항상 MOCK 제안 먼저 생성. 실주문 전환은 사용자 명시적 승인 필요 — 자동 실주문 없음.
- HMAC 서명 승인링크. 이메일/알림 링크에 TTL 포함 HMAC 토큰 삽입.
OrderProposalExpiryJob이 만료 큐 항목 자동 정리. - KIS OpenAPI (한국 주식) + Binance API (암호화폐). 두 거래소 모두 실주문 지원. 자격증명은 AES-GCM 암호화 저장 (
APP_CRYPTO_KEYenv var). 토큰 자동갱신, 잔고/시세/주문 API 완전 연결. - 글로벌 Kill-Switch.
TRADING_KILL_SWITCH=true설정 시 어댑터 레이어에서 모든 실주문 차단 — 환경변수 하나로 전체 라이브 트레이딩 즉시 정지.
Multi-LLM Daily Briefing (Gemini 2.5 · Claude) & 구독 인프라
- 멀티 LLM 일일 브리핑.
AiGatewayService가 Gemini 2.5 + Claude (Anthropic) 멀티 LLM 폴백 체인 관리. 일일 시장 요약 + 사용자별 포트폴리오 코멘트 자동 생성. OpenAI / Perplexity 라우팅도 지원. - Subscription plans. FREE / STANDARD (9,900₩/mo) / PREMIUM (19,900₩/mo) / EXPERT — Toss Payments v1 full payment flow. Sandbox testable without real billing.
- Rate limiting & security. Bucket4j enforces 20 AI chat req/hour/user. JWT HttpOnly cookies (XSS protection). Flyway migrations with
ddl-auto=validate. - 3-process EC2 deployment. Single
who-aEC2 instance: Nginx reverse proxy + 3 systemd units (Spring Boot :8080, FastAPI analytics :8001, React dist). Notification center via Zustand persist + BE/api/notifications/*.
🌉DevBridge - Freelancer × Client AI Matching Platform
Full-stack web application · Spring Boot 3 + React 19
A platform that matches freelancers and clients using natural language queries combined with structured filters. Powered by Google Gemini 2.0 Flash for AI-based match scoring, with a full project lifecycle from discovery to contract, milestone tracking, and escrow settlement.
Core Features
- AI Matching: Google Gemini 2.0 Flash scores partner/client/project compatibility from natural language search queries; falls back to rule-based scoring when API key is absent
- Search & Discovery: Partner, client, and project search with skill/field/budget/level filters; 1,000-record seeded dataset with N+1 optimized batch loading (3,001 → 4 queries)
- Project Lifecycle: Full application state machine — APPLIED → ACCEPTED/REJECTED → CONTRACTED → IN_PROGRESS → COMPLETED
- Milestone & Escrow: Milestone lifecycle (PENDING → IN_PROGRESS → SUBMITTED → APPROVED/REVISION_REQUESTED) with automatic escrow release on approval
- 7-Module Contract Negotiation: scope / deliverable / schedule / payment / revision / completion / terms — dual-accept pattern per module
- Stream Chat: Three room types — DM, NEGOTIATION, PROJECT_MEETING — via Stream Chat Java SDK
- Portfolio Management: Project-based or free-form portfolios with public/private visibility control
- Integrated Dashboard: Schedule (FullCalendar), applications, meetings, project management, evaluation, and portfolio tabs for both partner and client roles
- Bank Verification: 1-won mock auth with SecureRandom, 5-min TTL, 5-attempt limit, atomic ConcurrentHashMap processing
Technical Stack
Backend:
- Spring Boot 3.4.0 / Java 17: 22 REST controllers, layered architecture (controller → service → repository)
- JWT (HS256, 24h TTL): HttpOnly cookie delivery;
JwtAuthenticationFiltercookie-first with Authorization header fallback — XSS token theft blocked - MySQL 8 / JPA / Hibernate: 32-table schema, batch IN queries + JOIN FETCH for N+1 elimination
- Google Gemini 2.0 Flash: AI match scoring (rule-based fallback when key absent)
Frontend:
- React 19 (Vite): 38 routes, component-based SPA, React Error Boundary at root
- Zustand (persist): loginUser, interests, profileDetail state with localStorage persistence
- Axios + withCredentials: 17 resource-specific API adapters; 401 → auto redirect with SILENT_401_PATTERNS for UX preservation
Infrastructure:
- AWS EC2: systemd service management, production-grade
application-prod.properties(ddl-auto=validate, env-required secrets, actuator minimized) - Security hardening: path traversal prevention, COMPLETED-project mutation guard, CSRF via SameSite=Lax, multipart upload size cap (50 MB)
Scale & Performance
- 1,000 partners · 1,000 clients · 1,000 projects (seeded)
- Frontend display sample: 500 records per entity (API-layer
.slice(0, 500)) - Batch query optimization: 3,001 SQL → 4 SQL for partner list endpoint
- AI matching: up to 50 candidates per Gemini call
☀️Festory - Festival Travel Website 🌇
Frontend Development Project (React + Vite)
A modern web application that helps users discover festivals tailored to their preferences through AI-powered recommendations and comprehensive festival management features.
Core Features
- AI Taste Test: 7-question survey analyzes user preferences to recommend personalized festivals
- Google Calendar Integration: Manage festival schedules with full OAuth 2.0 integration
- Interactive Maps: Google Maps API integration for location visualization and nearby accommodation search
- Real-time Weather: OpenWeather API integration showing weather for 8 major cities
- Social Authentication: Google, Kakao OAuth login support
- Review System: Write and manage festival reviews with media uploads
Technical Stack
Frontend Framework & Build Tools:
- React 19.2.0: Latest version with functional components and Hooks
- Vite 7.2.4: Ultra-fast build tool with ES Modules and Hot Module Replacement
- React Router DOM 7.12.0: Client-side routing with BrowserRouter
Styling & UI:
- Tailwind CSS 4.1.18: Utility-first CSS framework with custom theme extensions
- Shadcn/ui Components: Radix UI-based accessible component library
- Lucide React 0.562.0: SVG-based icon library
State Management & Data:
- Zustand 5.0.10: Lightweight state management with localStorage persistence
- date-fns 4.1.0: Modern date utility library