Byun Seungjun 변승준

System Stability · Performance

• Root-cause incident fix — Ended a fleet-wide restart loop from connection-pool exhaustion by deriving “instances × pool cap ≤ DB limit” (4×10=40 / MySQL cap 151), stabilizing connections from 150+ to 40–50.
• Performance — Removed dashboard N+1 (lazy loading), consolidated main-list API calls to cut request volume ~66%, and rolled out a shared image-cache pipeline at ~45 call sites to reduce request and memory cost.
• Auth · Redis — Refresh-token rotation (RTR) with reuse detection and bulk revocation, distributed locks, and token blacklisting; concurrent 401 refreshes are serialized through a single client-side queue.
• Security by structure — Isolated external API keys behind a server proxy (path whitelist, Redis TTL cache); moved all credentials into GitHub Actions Secrets — “structure, not vigilance, prevents it.”

Architecture · Payments

• Two services, one backend — led the merge — Designed a single Spring Boot backend with unified JWT auth, and led the server merge with two incoming engineers by agreeing on data/transaction boundaries first.
• Toss payment infrastructure (recurring / one-time) — Amounts are server-computed, webhooks are double-verified (signature + PG re-query), and deterministic idempotent order keys block double charging.

Infra · Launch

• Deployment pipeline — GitHub Actions CI/CD with Blue-Green zero-downtime deploys on AWS (EC2·ALB·RDS·ElastiCache).
• App launch · content automation — Shipped the Flutter app to the App Store and Google Play (through repeated Apple review), and widened reach with an admin tool that turns data visualizations into card news.

Field · Collaboration

• Co-ran the CAMP-US Cup with the Dongducheon Football Association — same-day incident response on match day. Established PR / commit / issue templates for the team.

“Deliberate 50% less, execute 10% more (50210)” — I'd rather build small and learn fast than polish a perfect plan. The less precedent a problem has, the more I set my own hypothesis and bar, then let real users and data set the direction.

• Test the most uncertain assumption first, in the smallest way — and judge on metrics with visible cause and effect, not vanity numbers. Performance gets cut from real measurements; incidents get driven to their root cause so they can't recur.
• Remove repetition through structure and automation. I run an AI agent (Claude Code) in parallel sessions — build · code review · security audit — and keep an ADR (architecture decision record) practice so decisions and their context become team assets.
• Tools get delegated, judgment doesn't — only well-defined tasks go to agents; spec definition, output review, and the final call stay with me.