Seongju Kang강성주

2025.03MWC25 GLOMO Award: Best Mobile Innovation for Enhancing the Lives of Children and Young PeopleGSMA · SKT × YATAV MetaForestMy part 메타포레스트 Unity 클라이언트 개발

2024.112024 Metaverse Innovation Award: Excellence Prize과학기술정보통신부 주최 · 머니투데이·한국메타버스산업협회 주관

2023.11KT "Warm Tech Plus" top company: Minister of SMEs and Startups AwardKT · 중소벤처기업부

• 2026.06AEGIS: GS Certification Grade 1 · TTA · My part: GS 인증 결함 1·2·3차 대응
• 2025.09Selected for the Ministry of SMEs 'Hyper-gap Startup 1000+' Micro DIPS program · 중소벤처기업부 · 과제: AI 상담사 트레이너
• 2024.108th Seoul Forest Social Venture Expo: Grand Prize, Scale-up division · 성동구 · AI 얼굴인식 + 메타버스 상담 솔루션 · My part: 수상 제품의 Unity 클라이언트 개발

Videos (company YouTube channel)

• 2025-03MetaForest at MWC2025 GLOMO · 3:00
• 2024MetaForest promo (Korean) · 0:57
• 2024MetaForest promo (English) · 0:57
• 2024New space demo: digital-addiction prevention room · 0:36

• 01I replace foundations in a live service.Unity 2021 → 6, Photon PUN → Fusion and voice Photon Voice → Vivox → Agora were swapped without cutting users off, because PC and mobile QA lists (77 items for the August 2023 launch) and an emergency-release routine were in place first.
• 02I build the real-time modules myself.Screen sharing (WebRTC with Photon RPC ownership), 1:1 video chat, a multi-user whiteboard, expression tracking, voice (Agora, Vivox) and the Photon CCU split: the real-time features were mine.
• 03I own the delivery path end to end.From a WPF installer/updater in 2021 to an Electron launcher with 6-way parallel CDN downloads in 2026, one person shipped and announced four platforms and store reviews: Slow downloads in South Africa and Taiwan were solved by swapping servers and storage.
• 04I fix causes, not symptoms.The voice dropouts were not Agora but Photon disconnects amplified into 30-second silences by our code; the launcher update failures were not the server but a missing HTTP cache invalidation. When a measurement changes the diagnosis, I say so.
• 05I work toward building together.Release decisions go through PC and mobile QA lists, release notes include known issues and workarounds so institution staff can act, and handovers leave a HANDOFF document the next person can pick up.

• 
  MetaForest Client: from launch to 3.x

  Four years building and shipping the metaverse counseling client: installer, screen sharing, video chat, whiteboard, transcripts, and the Vivox migration

  Unity client development · production releases · Unity · 2021 – 2025

  2021.10 – 2025.12
• 
  MetaForest Launcher

  Built a Windows and macOS desktop launcher from scratch and moved distribution to a CDN

  Design, implementation, deployment · Electron · Windows / macOS

  2026.06 – 2026.08
• 
  MetaForest Unity Client

  A Unity 6 remake of the client, run side by side on PC and mobile in the same room

  Client development · Unity · Photon · Agora

  2025.11 – 2026.08
• 
  AEGIS: AI Security Gateway (AI DLP)

  QA-driven development and rollout on a product that judges and blocks prompts headed for external AI services

  QA-driven development and rollout · Certification test defects · AI DLP rollout in the financial sector

  2026.06 – present
• 
  Independent Intelligence: Air-gapped Workplace AI

  Pre-launch QA, gateway fixes, a full server rebuild and operations for an on-premise AI appliance

  QA · gateway development · server rebuild and operations · LibreChat · LiteLLM gateway hooks · vLLM / Ollama · Jetson

  2026.07 – present

Other work: MetaForest Web & Backend, Cloud Cost Reduction · Cost Dashboard, YATAV Company Blog, University AAC Chatbot Backend, NICE Identity Verification S2S Migration

Writing and verifying code with AI tools is daily work. Below are things I can show as artifacts rather than claims. The code is taken from the repositories as is (comments are in Korean).

1. 01Built my own agent: seongjuclaw

   Python · Claude API · Ollama · Jul 2026 –

   What

   An agent that uses the terminal, files and the web as hands. Tool loop, a SKILL.md skill system, session save/restore, serverless local RAG (fastembed), local models via Ollama, a closed-network mode that blocks every outbound call, a desktop GUI.

   Why

   To direct tools like Claude Code well, I wanted to know how they run inside: tool-call ordering, context management, the safety conditions for parallel execution.

   Verified

   RAG and the tool loop were run end to end with a real model. The GUI was run in closed-network mode with a local model (gemma4 12b) reading files and answering, captured in September 2026.

   def _run_tools(self, tool_uses):
       """한 턴에 요청된 도구들을 실행.
       연속으로 붙어 있는 '병렬 안전'(tools.PARALLEL_SAFE) 도구는 묶어서 동시에 돌린다 —
       파일 여러 개 읽기·grep 여러 번 같은 탐색이 대부분이라 체감이 크게 빨라진다.
       쓰기·부작용 도구를 만나면 거기서 끊어 순서를 지킨다: read→edit→read 처럼
       앞뒤 순서가 의미를 갖는 요청이 뒤집히면 안 되기 때문."""
       n = len(tool_uses); results = [None] * n; i = 0
       while i < n:
           j = i
           while j < n and tool_uses[j].name in tools_mod.PARALLEL_SAFE:
               j += 1
           if j - i < 2:  # 병렬 이득이 없으면(0~1개) 그냥 순차
               results[i] = self._one_tool(tool_uses[i]); i += 1; continue
           with cf.ThreadPoolExecutor(max_workers=min(4, j - i)) as ex:
               futs = {ex.submit(self._one_tool, tool_uses[k]): k for k in range(i, j)}
               for f in cf.as_completed(futs):
                   results[futs[f]] = f.result()
           i = j
       return results

2. 02Moved work to local models to cut the token bill

   Ollama · llama.cpp · vLLM · fastembed · 2026

   Why

   Every extra API call cost money. Research, summarizing and classification do not need top-tier models, so they moved down to local ones.

   How

   My own agent runs Ollama on the same tool loop as Claude, so only the model swaps, and a closed-network mode blocks every outbound call. RAG runs serverless on fastembed and numpy, so no embedding API either.

   At work

   The air-gapped AI product runs entirely on local models (vLLM, Ollama). Moving embeddings to the GPU cut the first response from 20 s to 1.2 s, and the AI security product runs its judgment model locally on llama.cpp.

   Limits

   A local 9B loses formatting on long context, so I split work by task between local and API.

3. 03Built Unity-specific build automation

   Python · Unity batch mode · Aug 2026

   Why

   Five years of CI kept tripping on the same things: exit code 0 on a failed build, -quit conflicting with test options, license seats never returned.

   How

   I wrote the traps you only learn by getting burned as module documentation first, then let the agent implement. Domain knowledge from me, implementation speed from AI.

   How far

   The verdict logic that trusts the build report and log signals over the exit code is pinned by tests. This Mac does not have that Unity version, so it was never run end to end on a real project, and I say so.

   def verdict(summary, returncode, *, report_result=None):
       """실행이 성공인지 최종 판정한다.
       우선순위는 빌드 리포트 > 로그의 batchmode 신호 > 종료코드 순이다."""
       if report_result is not None:
           if report_result != "Succeeded":
               return False, f"빌드 리포트 결과가 {report_result}"
           if summary.errors:
               return False, f"빌드는 성공했지만 로그에 에러 {len(summary.errors)}건"
           return True, "빌드 리포트 Succeeded"
       if summary.batchmode_result == "failure":
           return False, summary.abort_reason or "배치모드가 실패로 종료"
       if returncode != 0:
           return False, f"종료코드 {returncode}"
       if summary.errors:
           # 여기가 핵심: 종료코드는 0인데 로그엔 에러가 있는 상태.
           return False, f"종료코드는 0이지만 로그에 에러 {len(summary.errors)}건 (배치모드 종료코드 오탐)"
       return True, "정상 종료"

4. 04QA, build and operate AI products

   AEGIS · Independent Intelligence · AAC chatbot · 2026

   AEGIS

   An AI security gateway (AI DLP). GS certification defect follow-ups (one defect = one PR); for a financial-sector rollout, a pre-delivery checklist measured by hand surfaced role-boundary, audit-ledger and Korean PII detection defects, which I fixed.

   Independent Intelligence

   An air-gapped workplace AI. Pre-launch QA, gateway hook and document tool fixes (0% → 99% over 100 runs), and a rebuild of the dead server on a Jetson with two domains live.

   Rule

   Agents may write code but never deploy. Irreversible actions are pressed by a person.

   ---
   name: code-review
   description: 코드나 변경사항(diff)을 리뷰할 때. 버그·정확성 위주로 살펴보고 우선순위를 매겨 보고합니다.
   ---
   1. 대상 파악: 특정 파일이면 read_file, 변경사항이면 run_bash로 git diff를 확인합니다.
   2. 정확성(경계 조건·null·off-by-one·예외) → 리소스 → 동시성 → 보안 → 단순화 순으로 봅니다.
   3. 발견 항목을 심각도 높은 순으로: [심각도] 파일:줄 — 요약 / 재현 시나리오 / 제안 수정
   주의: 확실하지 않은 항목은 "추정"이라고 밝히세요. 스타일 문제는 마지막에 짧게.

When I hand work to AI

• Rules go in documents. CLAUDE.md and HANDOFF.md per repo, SKILL.md for the agent. What not to do comes first.
• Domain knowledge from me, implementation together. Traps you only learn by getting burned are written down first. Not knowing that a dev auth bypass silently falls back to admin makes you read a permission test backwards.
• Results get re-checked. Build, run, screenshot, measure. "Done" only after I have seen it.
• Irreversible actions stay human. Deploy, publish, delete, pay: never the agent.

• 
  PlanEatJJJ Studio · team of three

  A mobile survival game where you grow by swallowing space junk up to planets: team of three, preparing a Google Play release

  Unity 6 · C# · URP · UGS

  2026
• 
  girlsXzombies

  A 2021 tower-defense game project

  Unity · C# · HLSL

  2021
• 
  SeongjuClaw성주클루

  An AI agent desktop app that runs Claude and local models on the same tool loop, with a closed-network mode

  Python · pywebview · Claude API · Ollama

  2026

2026

A financial-sector AI DLP rollout and GS certification defect work on AEGIS, operating and rebuilding an on-premise LLM chat service, the MetaForest Electron launcher and CDN migration, the NICE S2S identity migration, a cost dashboard. Personal: the three-person mobile game PlanEat (Google Play release prep) and the seongjuclaw agent.

2025

MetaForest 3.0 overhaul, Unity 6 and Photon Fusion migration, Photon Voice → Vivox, AI summaries for transcripts. MWC25 GLOMO Award.

2024

Whiteboard with real-time sync, STT transcripts and the move to Naver Clova, social login, deep links and an installer renewal, MWC 2025 preparation with SKT. Metaverse Innovation Award.

2023

Official launch with a Seoul Wee-center pilot, the August store release on iOS and Android, avatar customization, video chat, a GPT-based AI NPC demo for CES 2024.

2022

Screen-sharing module, 0.1.0 and 0.5.0 updates, per-institution trial builds, the mobile version and Addressables, the first mobile production release.

2021

Joined in July. First institutional deployment in October, a custom installer and updater, sprint-1 pilot QA.

Full timeline (84)

Unity · Game

Unity 2021 LTS → 6C#URPNew Input SystemAddressablesPhoton PUN · Fusion · VoiceVivoxAgoraMediaPipePlastic SCM

Platform

Windows · macOS · Android · iOS 빌드·스토어 심사WPF · Electron 런처S3 · CloudFront 배포GitHub Actions

AI

GPT · Claude APIWhisperNaver Clova STTOllama · vLLMRAG에이전트 도구 루프

Languages

C#PythonTypeScriptKotlinSwiftDartSQL

Backend · Infra

Node.jsMySQL · PostgreSQLWebSocket · WebRTCAWS EC2 · S3 · LambdaCloudflare WorkersDockernginx