SangHyun Bang 방상현

Backend Developer

• ERP accounting module · Node.js, MariaDB

Student Researcher

• Assisted Multi-Armed Bandit / reinforcement-learning research — theory analysis, experiment organization, project support

Research Intern

• Bandit Algorithm / Reinforcement Learning research intern

Management Planning Intern

• Supported management-planning work — experienced how business and data connect to decisions

• Solo — planning and design docs (23 ADRs), Flutter app, Spring Boot server, AWS infrastructure (Terraform), ONE store release (2026.09) and operation; about 81% of commits written with an AI coding agent (marked by trailers)
• Gathered the nine places where coordinates can leave the device into one compile-time policy, and computed nearby places on the device from a nationwide catalog (no location-based service business filing has been made)
• Chat: shared rooms for family and crews, HTTP send with ETag polling (3→30 s), per-room sequence issued atomically with LAST_INSERT_ID(last_seq + 1) and pinned by a 20-thread test on real MySQL
• Single-EC2 production (Caddy as the only entry, DB data on separate EBS, config in SSM, CI runs tests and doc-drift checks, manual deploy). After a Caddy config error on first boot, replaced the string check with a test that runs caddy validate in the production image · 1,994 backend tests, 0 failures (2026-09-30)

A valuation service that estimates a fair-value range with public valuation methods (DCF · Monte Carlo), written with an AI coding agent (co-authorship marked in commits). Pure Kotlin engine + Spring Boot API (PostgreSQL · Testcontainers) + Flutter app. A private project being prepared as a business; the source is not public. Not investment advice.

• Owned the auth boundary of a five-person Spring Cloud MSA — member/auth service, gateway auth filters, OPA policy and bundle service
• Gateway validates the cookie JWT and rewrites user headers; authorization is an OPA query over per-service, per-role path rules, failing closed with 503 on OPA errors or delay (circuit breaker, bulkhead, 2 s)
• Designed and built the path that publishes suspensions, withdrawals and seller approvals as Kafka hotlist events; a bundle service rebuilds the policy bundle and OPA pulls it every 10–60 s; the seller service syncs from the same topic
• HttpOnly/Secure/SameSite=Strict cookies, refresh-token rotation and revocation, Kakao/Naver OAuth account linking, a withdrawal outbox (written in the same transaction, FAILED after 5 attempts)

🏆 Excellence Award, 2025 Tourism Data Utilization Competition (Korea Tourism Organization, 2025.11)

• Team lead (planning) of a three-person contest team — backend (place collection, explore, recommendation, media, auth), Python recommendation server, app recommendation & explore screens, deployment and ONE store release
• Gave the LLM 95 category names instead of places (two groups chosen); Spring maps labels to cat3 codes, filters places and ranks them with its own scoring
• Collected and cleaned ~50,000 Tourism Organization places, then backfilled Kakao links & coordinates and built category trait snapshots; infinite scroll uses a (createdtime, contentid) cursor with size+1
• (2026, solo) Prepared for public operation with auth on the recommendation API, per-path rate limits and secret scanning; for search, found the count query's full scan with EXPLAIN ANALYZE and switched to search_text + FULLTEXT (ngram) (count ~755 ms before, list ~4.9 ms after — different queries, single measurement)

🏆 Excellence Award, KDT Full-stack Bootcamp graduation project (Gangnam HiMedia, 2025.07)

• Set up the initial structure of both backend and frontend repositories (Spring Boot; React routes, Redux, dev proxy) and managed server configuration and build files
• Backend and frontend for the map (mountain search · mountain forecast · trails), curation posts, social login (OAuth2 · JWT) and S3 upload (Spring Boot · React)
• Forecasts for 153 mountains collected in batches and kept in Redis (12-hour TTL); the map’s forecast API reads only from Redis (the automatic refresh trigger is off in the final code)
• Curation posts keep post metadata in MySQL and route segments (GeoJSON coordinates) in MongoDB, linked by postId and drawn as a route on the map