윤형주

Stack Java 8 · Spring MVC 3.x · eGovFrame 4.3 · Tibero · Jenkins  |  hrd.rda.go.kr/ehrd_front

배경A full rebuild of an aging public-sector LMS, where meeting security and web-accessibility standards was the precondition to launch.

역할Owned the backend for core user/admin features and Tibero query tuning.

• Developed 51% of core user/admin features — member management, boards, messages, admin dashboard, enrollment
• Improved initial load time 58% via Tibero index tuning and subquery refactoring: 5.7s → 2.4s
• Passed the web vulnerability inspection (SQL injection, XSS, and more) and the KWCAG-based web accessibility inspection, securing public-launch quality
• Coordinated requirements directly with the client and reflected audit bodies' findings within schedule

Stack Spring Boot 3.4.5 · Java 21 · FCM · JPA · MyBatis · JWT · Spring Security

배경Each service reimplemented push, AlimTalk, and SMS separately; a unified server several teams could share was needed.

역할Owned the design and implementation of the plug-in server, send-performance optimization, and a stateless auth scheme.

• Designed and built a plug-in notification server unifying push, AlimTalk, and SMS
• Cut 1,000-recipient send time from 30s to 5s via async processing and batched FCM calls
• Improved history-search time from 3s to 0.5s with dynamic queries and composite indexes
• Built a stateless auth scheme with JWT Access/Refresh tokens, BCrypt, and device-ID-based refresh
• Documented a shared-module integration guide and supported adoption across service teams

Stack Python · Django · Llama · Qdrant · vLLM · LangGraph · Docker · NVIDIA CUDA 12.3

배경Staff struggled to find answers across vast internal manuals, and the closed network ruled out external LLM APIs, so an on-prem RAG system was needed.

역할Implemented AI/backend end to end — from the embedding pipeline to LangGraph retrieval/grading graphs and vLLM on-prem serving.

• Designed an embedding pipeline for PDF, HWPX, Excel, and PPT and loaded it into a Qdrant HNSW vector DB
• Implemented a LangGraph-based Rewrite, Retrieve, Grade, Regenerate RAG pipeline
• Lifted accuracy ~40% over keyword search via LLM-as-Judge relevance grading and auto re-retrieval
• Built vLLM on-prem serving in the closed network, reaching 2–3s average response with no external API dependency
• Verified accuracy against the field team's manuals and QA set and shared results to prove adoption value

Stack Spring Boot 3.4.5 · Playwright · Selenium · BeautifulSoup4 · MySQL · Qdrant  |  jdcc.or.kr/policy

배경The goal was to automate the manual checking of projects, bids, and notices scattered across many agencies, and gather them in an LLM-analyzable form.

역할Designed per-agency crawlers and collection agents and built the backend pipeline that refines and structures the data.

• Automated collection of projects, bids, and notices across 7 agencies — NIA, NIPA, IITP, KIRIA, KIAT, KOSME, TIPA
• Built per-site crawlers and collection agents to ingest policy/R&D data in a structured form
• Implemented a backend pipeline that refines the collected data into an LLM-analyzable form

Stack Spring · Oracle · Nexacro · eGovFrame 4.3 · Jira · Confluence

배경Not one project but several university and public-sector systems at once — incident response, performance, and certification in parallel.

역할Handled operations overall — maintenance, performance tuning, security checks, GS certification, and new-server setup.

• Enhanced ~28% of Kunsan Univ. academic/faculty systems; improved performance with MyBatis mappers and index tuning
• Ran SSL/TLS, browser-compatibility, firewall, and packet checks on the National Pension Service insurance linkage center
• Designed, built, and maintained Oracle tables for the Gunjang Univ. scholarship system with customer support
• Maintained Howon Univ. academic/faculty/main systems with customer support and scheduled server checks
• Earned GS Certification Level 1 for in-house CMS/LMS via transaction-integrity fixes and Maven dependency resolution
• Documented 10 in-house servers and set up OS/security/network for 3 new ones
• Blocked 12,000+ SSH brute-force attempts (Fail2Ban, port change, key auth) and handled a mail-server security incident
• Managed issues and deliverables in Jira/Confluence and ran customer training and inspection reporting

howon.ac.kr/web/main.html

Stack FastAPI · LangGraph · Spring Boot 3.4.5 · Qdrant · Elasticsearch · Neo4j · Kakao Cloud · Kubernetes

배경Designed to go beyond rating-based recommendation toward natural-language intent recommendation, with Cold Start and single-retrieval bias as the core problems.

역할As PM and backend lead of four, led the hybrid RAG recommendation engine, the payment/reward system, and API design and scheduling.

• Built multi-agent routing over LangGraph (5 graphs / 52 nodes), splitting recommendation flows by intent
• Implemented a hybrid RAG engine running vector, BM25, and graph search in parallel and fusing them with RRF
• Used content-based recommendation for Cold Start, switching to collaborative filtering and diversity rerank as data accrued
• Implemented the subscription lifecycle: idempotency-key duplicate-payment block, batch cleanup of unpaid orders, billing-key auto re-charge
• Auto-matched 55 activity-reward types with concurrency control using pessimistic locks and AOP
• Processed 1.17M movies / 58GB via 2,000-row streaming with checkpoint resume, resolving OOM
• As PM and backend lead of four, owned API design, scheduling/role coordination, and code/doc standards

User app·Admin·GitHub·AI Agent doc·Backend doc·Design doc

Stack Spring Boot 3.4.5 · Spring WebFlux · Java 21 · MySQL · Redis · Elasticsearch · Docker · AWS · React Native

배경A finance app I ship and operate myself. Couples, families, and groups manage one shared ledger together, with an offline-first experience that works even without a network.

역할As a solo developer, carried out the entire process alone — server, app, web, and infrastructure.

• Unified personal and shared ledgers around a Book-centric domain model; built a double-entry engine with journaling, account trees, and auto-generated financial statements
• Redis caching cut the statistics API from 2s to 50ms (40×); external calls on WebFlux non-blocking I/O, internal logic on Virtual Threads
• Offline-first — every feature works without login; on login, local data syncs in transactions (rollback on failure) with local→server ID mapping
• Implemented Elasticsearch Korean full-text search and an audit trail for query and traceability
• After a key-leak incident, redesigned the deploy path to bypass any registry; an instance downsize cut running costs ~50%
• Operational visibility with ELK and Prometheus/Grafana; sole operator from design to incident response since launch
• Social login with 4 OAuth2 providers (Google·Naver·Kakao·Apple) and RS256 asymmetric-key JWT — JWKS/kid-based zero-downtime key rotation and jti replay detection against token theft

garabu.org·App Store·Google Play·Server doc·AI Agent doc·App doc·Web doc

Stack Spring Boot 3.4.5 · MySQL · Redis · AWS · React Native · Expo · Docker

배경Aiming beyond solo workout tracking toward a community that shares and cheers in real time, I designed the notification and real-time transport from scratch.

역할As a solo developer, implemented the REST API, real-time transport, social login, and the deploy pipeline.

• Designed 130+ REST endpoints across 6 domains with SpringDoc OpenAPI auto-docs
• Implemented JWT and OAuth2 (Google, Apple, Kakao), a Spring Security filter chain, BCrypt, and rate limiting
• Built SSE-based real-time alerts/workout broadcasts with Redis Pub/Sub multi-instance broadcasting
• Set up the real-time alert flow to auto-fall back to FCM push for disconnected users
• Built the deploy pipeline on AWS EC2, RDS, S3, SES with Docker multi-stage builds

App Store·Google Play·Server doc·App doc