박상욱Steve Park

Genre: Social Networking Service (SNS) · Cozy Life Simulation

• Login bottleneck fix (30s–3min → 10–30s) · Ranking Redis migration (10–20s → 1–2s) · New chat server · Channel distribution logic

• Stabilized large-scale user environment through data management and optimization→ While developing on the CocoMine Sanrio codebase, identified bottlenecks and applied indexing/query optimization, reducing item and ranking data processing from 5–15s to 2–6s. Reworked the login system — narrowing per-API locks to per-user scope so different users' logins run in parallel, backed by ConcurrentDictionary-based concurrent structures — cutting login time from 30s–3min to 10–30s.
• Developed chat/plaza/village/district servers for diverse user interactions→ Added village and district servers for new content variety; introduced interactive items, interaction motions, and an address system within the village.
• Redis-based ranking management and game server content development→ Migrated the frequently-queried ranking system to Redis, reducing query time from 10–20s to 1–2s.
• Team collaboration and technical direction review→ Worked as one of two server developers in a team of 2 server · 5 client · 1 UI, handling server-side content development. Coordinated implementation schedules and test plans in team meetings, and reviewed key technical decisions — such as channel distribution and scale-out logic — with the CTO before applying. After the other server developer left mid-project, continued solo and following launch, independently identified and resolved structural issues in the existing codebase.
• Pre-launch load testing via custom bot — bottleneck detection→ Built an automated load-testing bot covering connection, login, and basic gameplay to verify stability before launch. Identified high load on the login, item, and plaza/chat paths, then restructured those areas and reviewed the codebase for lock contention, race conditions, and potential deadlocks.
• On-call incident response — missing event launch · login delays→ Responded to a late-night incident where a planning team event failed to go live; investigation revealed the data had been omitted from the commit. Received the re-committed file and hot-deployed without downtime to restore the event. Also resolved a separate incident where a specific user's excessive item count triggered a lock that delayed all other users' logins — applied indexing and a limit-based mitigation immediately, then redesigned the lock structure as a long-term fix.
• Production data mismatch response and prevention — achievement system→ After adding achievement features and passing staging tests, a data mismatch surfaced in production for users who had already sold or consumed the relevant items, leaving their achievements unprocessable. Analyzed usage logs and transaction history to identify affected users and applied a bulk achievement update to restore correct state. Added more granular edge-case checks to the server-side achievement ingestion logic to prevent recurrence.

Genre: Social Networking Service (SNS) · Metaverse Community

• Led full administrative system development; improved server performance through data management and optimization
• Built chat and plaza servers for real-time user interaction
• Redis-based ranking management and thread-based session isolation for stability and scalability

• Led full administrative system development; improved server performance→ Designed and built the ASP.NET Core-based administrative page for the operations team, added a BigQuery-powered game currency and item statistics dashboard, and removed redundant or overlapping logic to improve performance.
• Built chat and plaza servers for real-time user interaction→ Extended the solo farming and decorating space into a shared environment where users can chat and see each other's avatars.
• Redis-based ranking management and thread-based session isolation→ Migrated the ranking system to Redis, reducing query time from 10–20s to 1–2s. Also isolated user session handling into separate threads so a delayed request from one user no longer blocks other users' sessions, keeping the server responsive as concurrency grows.

Genre: Styling Game (Fashion Dress-Up)

• Server maintenance and administrative system feature additions to improve operational efficiency
• Data management and optimization for stable large-scale user data processing
• Performed DB migration from AWS to on-premise server

• Server maintenance and administrative system feature additions→ Tracked intermittent errors, resolved conflicts between legacy code and new content, and added a statistics system and new content management features to the administrative system to help the operations team manage the game more efficiently.
• Data management and optimization→ Improved poorly optimized queries and missing indexes, reducing stage-clear item delivery processing time from 10–20s to 2–5s.
• DB migration from AWS to on-premise server→ Performed large-scale data backup and migration as part of a cost-reduction move from AWS to on-premise.

Genre: Match-3 Puzzle Game

• Added administrative system features to support efficient game management by the operations team
• Data management and optimization for stable game data processing
• Performed DB migration from AWS to on-premise server

• Added administrative system features for efficient game management→ Added a currency flow statistics system to help the operations team track how in-game currency is used and distributed.
• Data management and optimization→ Consolidated and cleaned up multiple DB columns during a game system overhaul; deployed after thorough testing with no disruption to gameplay.
• DB migration from AWS to on-premise server→ Performed large-scale data backup and migration as part of a cost-reduction move from AWS to on-premise.

Genre: SNG · Lifestyle Simulation

• Payment failure rate improvement (4–5/day → 1/1–2 weeks) · MySQL 5.7→8.0 migration · Looker Studio dashboard build
• KR/JP/TW/EN 4-region server maintenance

• Stable service operation across KR/JP/TW/EN 4-region servers→ Tracked down and fixed errors caused by differing content per region; created and continuously updated a content/bug documentation list so other developers could also manage it smoothly.
• Server performance improvement and new content development through data management and optimization→ Applied indexing and query optimization during new content development to ensure features like item systems worked correctly without data duplication.
• MySQL version migration — source code fixes to secure system compatibility and stability→ With MySQL 5.7 reaching end-of-life, an upgrade to 8.0 was required. Identified breaking changes (authentication method, GROUP BY handling, etc.), updated source code, ran thorough testing, and deployed with zero issues.
• Looker Studio data visualization dashboard to improve operations team efficiency→ Connected BigQuery with Looker Studio so game currency and log data could be checked at a glance on the dashboard, directly supporting the operations team's decision-making.
• On-call incident response — Japan server disk-full outage→ Responded to a late-night outage on the Japan server. Running df confirmed the disk was full due to accumulated log files — the crontab-based log cleanup batch had been malfunctioning. Cleared old log files to restore the server, then audited and re-automated the batch job to prevent recurrence.

• Implemented administrative pages for both administrators and end users
• Integrated APIs from the education office, integrated reservation system, and booking system
• DB design and push-based notification system development

• Implemented administrative pages for administrators and end users→ Designed and developed separate administrative and user-facing pages with corresponding DB schema based on planning documents.
• API integration for data flow→ Called the existing integrated reservation system API from the education office and connected it to the platform for easy parent access.
• DB design and push notification system→ Designed the DB and built a push notification system to deliver new announcements and alerts directly to users' phones.

• Developed satisfaction survey system and school self-assessment features
• Fixed runtime errors to improve system stability

• Satisfaction survey and self-assessment features→ Added pages for parents and teachers to create satisfaction surveys, collect results, generate statistics, and export to PDF for administrative use.
• Error fixes for system stability→ Fixed broken or error-prone parts of the existing system, reducing user complaints.

• Solo development: DB design and full administrative page development
• OKTA SSO & KNOX MAIL API integration for security and usability

• Solo development: DB design and full administrative page development→ Independently developed the backend (excluding design and frontend), designed the DB for both administrative and user-facing pages.
• OKTA SSO & KNOX MAIL API integration→ Integrated Samsung Biologics' required SSO and KNOX MAIL, designed internal/external access separation, and coordinated L4 routing rules with the network team.

• Led full development as project PM
• Implemented resource management system and QR code recognition
• PWA-based system for mobile resource management

• Led full development as project PM→ Oversaw everything from planning to backend development (excluding design and frontend). Coordinated features and timelines with PK Valve staff.
• QR code recognition for resource management→ Built a mobile QR scan feature to instantly retrieve equipment information, making asset management more efficient.
• PWA-based system for mobile access→ Built as a PWA so users can access it like a native app on mobile without downloading from an app store.

• Replaced manual paper/Excel scoring for 6,400–10,000 staff transfers with score-based automated algorithm · tie-breaking logic
• Developed HR records, performance evaluation, survey, and statistics features

• Replaced manual paper/Excel scoring for 6,400–10,000 staff transfers with score-based automated algorithm→ Previously each person had to manually score and consolidate data through many steps. By automating the process, personnel transfer decisions became much faster. There was no longer a need to verify the process every time a score was assigned.
• Developed HR records, performance evaluation, survey, and statistics features→ The statistics feature made it possible to see which departments have the most transfers and which are most preferred. HR records that were previously stored in physical documents were consolidated into a single system for instant lookup.

• Developed favorites feature and user-facing pages
• Developed announcements and 1:1 inquiry features

• Favorites and user-facing page development→ Developed the community page favorites feature and portions of the user-facing content pages.
• Announcements and 1:1 inquiry features→ Developed an announcements page and 1:1 inquiry system within a tight deadline, reducing the burden on the overall site development.

• Solo development: video conferencing system and DB design
• TURN server setup
• Developed participant focus analysis algorithm using eye tracking

• Solo development: video conferencing system and DB design→ Built a remote communication system in Node.js enabling teachers and students to use chat, video, and audio like a real classroom. Designed the DB for chat history and access logs.
• TURN server setup→ Used WebRTC and ICE framework for direct connection attempts; on failure, routed through a TURN server. Repeated failures were treated as network latency and handled as connection failures.
• Focus analysis algorithm via eye tracking→ Used Mediapipe to track eye positions, measuring deviation from baseline gaze as a median over intervals. Sustained deviation beyond a threshold reduced the focus score.

• Operated video consultation and interview system for job seekers, interviewers, and counselors
• Adapted existing video conferencing solution for job fair environment

• Video system for job seekers, interviewers, and counselors→ Restructured BetterOn's classroom-oriented design for interviewer/interviewee and job seeker/counselor flows; added file attachment for document sharing.
• Adapted for job fair environment→ Logged all chat history and participant records for institutional requests. Added a mirror preview before entering sessions and a Mediapipe-based posing feature to help participants feel at ease.

Ramen Recipe and Subscription Web Service

• Developed user main page, subscription and gifting features
• Developed FAQ feature for customer support efficiency

• User main page, subscription and gifting→ Implemented the home page, subscription feature, and a gifting system allowing subscribers to send gifts to each other.
• FAQ feature for customer support→ Built an FAQ and inquiry page as a customer communication channel.

• Developed full administrative page for store and order management
• Developed user-facing app for intuitive ordering and delivery experience

• Administrative page for store and order management→ First project at the company. Built administrative features for managing store menus and information, and monitoring order status.
• User-facing app for ordering and delivery→ Developed in React Native with live previews; built order history and delivery status tracking for users.

KakaoTech Bootcamp Deep Dive · GPT Architecture from Scratch + Qwen3 Local LLM + RAG / LangChain / LangGraph / Claude API

• Designed GPT (~97M params) & jamo-unit BPE tokenizer from scratch; completed multi-stage training (continuation → QA → span → DPO)
• BM25+FAISS hybrid RAG + LangGraph 4-graph pipeline / Qwen3-1.7B local LLM in BF16 & GGUF Q4_K_M modes
• Claude Haiku auto question classification routing + 4-panel real-time comparison UI, FastAPI + SSE streaming

• GPT (~97M params) & jamo-unit BPE tokenizer from scratch + multi-stage training→ Implemented a tokenizer that decomposes Korean via NFD into jamo units and applies BPE. Designed a GPT decoder (CausalSelfAttention + MLP(GELU) + Pre-norm) from scratch, and built a staged training pipeline on kowikitext + AI Hub data: continuation → QA fine-tuning → span extraction → DPO preference learning, with bf16 mixed precision and cosine LR decay.
• BM25+FAISS hybrid RAG + LangGraph 4-graph pipeline / Qwen3-1.7B local LLM in two modes→ Upgraded from TF-IDF to a BM25+FAISS hybrid retriever, improving routing accuracy from 73.3% to 82.7%. Built 4 LangGraph StateGraphs (SOP_GPT / Claude / Qwen BF16 / Qwen Q4) using a node factory pattern. Integrated Qwen3-1.7B as both BF16 via Transformers+MPS and GGUF Q4_K_M via llama-cpp — benchmarked on 100 KorQuAD questions, Q4_K_M was 3.3× faster with <2%p accuracy gap.
• Claude Haiku auto routing + 4-panel comparison UI + FastAPI SSE streaming→ Claude Haiku classifies each question into casual / factual / general and routes it to the appropriate chain automatically. All 4 models (SOP_GPT · Claude · Qwen BF16 · Qwen Q4) stream responses simultaneously in a 4-panel UI via FastAPI + SSE. Login secured with PBKDF2-HMAC-SHA256; conversation context persisted across sessions via LangGraph MemorySaver (short-term) + JSON files (long-term).