Projects

In-depth write-ups of production services and larger projects.

Featured projects

Backend / Infra (led a 3-person team) · 2026.02 - 진행 중

  • A production Online Judge used in PNU’s mandatory course (about 800 users). Owned infra & backend and led a 3-person team.
  • Designed and built a local-LLM (vLLM) AI tutor, tuned via Locust load tests to serve 30 concurrent users within 3 seconds.
  • Ran the K3s cluster directly, diagnosing production incidents (crash-loops, a CNPG WAL runaway) down to root cause.

Full-stack (planning, development, operations) · 2026.02 - 진행 중

  • Consolidated scattered spreadsheets - company data, internships, industry outcomes - into one web system, now used in production (aicms.pusan.ac.kr).
  • Built solo end to end on Next.js and PostgreSQL: data modeling, features, and deployment.
  • Migrated from cloud (Vercel + Supabase) to an on-campus self-hosted server with Caddy HTTPS and GitHub Actions auto-deploy.

Solo architect & full-stack developer · 2026.05 -

  • Designed and built solo a zero-install CMS where non-developer staff create print-grade brochures with one line of Markdown, auto-deployed on save.
  • Implemented auth, collaboration locking, and CI/CD using only the GitHub API and git commits - no backend, no DB.
  • Prevented concurrent-edit conflicts with a distributed Soft Lock built on .locks commits.

More projects

Development (backend, app, deployment) · 2025.05 - 2025.08

  • A mental-health app that analyzes journal emotions with GPT (Plutchik’s wheel) and links students to campus counseling when negative emotions persist for 7 days. Encouragement Award, 6th PNU SW Hackathon.
  • As the team’s only developer, I built and deployed a FastAPI/MySQL backend and a Flutter cross-platform app end to end - my first app project after only doing web.
  • Designed a keyword fallback so emotion analysis keeps working when the LLM fails, a consent-based token-snapshot scheme for safely sharing 7-day summaries with counselors, and a layer-based renderer for the overlapping items in the "mind garden".

Backend · 2025.02

  • At a 2-week LikeLion backend bootcamp, solo-built the backend of a second-hand marketplace modeled on Karrot. (Frontend was a teammate’s.)
  • Built JWT auth, post CRUD, infinite scroll, and photo upload on FastAPI/SQLite/Redis, cutting list loading from 5-10s to under 1s with Redis caching.
  • My first project built with official docs only (no ChatGPT) - learned to debug by digging through docs and search.

RL (reward design, PPO training, experiments) · 2025

  • A capstone training a car to do perpendicular parking with PPO inside a Unity simulation. A 3-person team project at PNU.
  • I owned the RL side - framing parking as states/actions/rewards, training with PPO, and analyzing runs in TensorBoard. (A teammate built the Unity environment; another did the performance analysis.)
  • Got past the early cold-start with a demonstration file (imitation learning), then raised the estimated success rate from 75-90% to 95-99% by improving demo diversity and training budget across two experiments.