• Developed the frontend and landing pages for 5 services, including a Web3 DEX, a real-time game, and an AI agent.

• Real-time systems

  • WebSocket architecture: used an authentication-gateway pattern to guarantee connect-then-authenticate ordering, with event queuing to prevent message loss.

  • Connection reliability: exponential-backoff reconnection, plus Page Visibility API–based tab-inactivity detection to avoid holding unnecessary connections open.

  • Canvas-based chart rendering: smooth curves via Catmull-Rom spline interpolation, with DPR handling and frustum culling to sustain 60 fps.

  • Background-tab synchronization: buffered the data queue while a tab was inactive and batch-processed on return to restore chart state.

• Perpetual DEX trading (Orderly Network integration)

  • Multi-socket optimization: consolidated 4 independent sockets (chart / price / market / LP) under a singleton SocketService, reusing instances by URL + path to cut connection count by 50%, then diagnosed and resolved the underlying infrastructure issue.

  • 1-click trading: generated delegate accounts via EIP-712 signatures, enabling instant orders without a wallet approval on every trade.

  • TradingView chart customization: real-time position-line overlays (entry / TP / liquidation), visualizing entry price, liquidation price, and PnL.

  • Multi-chain support: branched contract addresses per network (Arbitrum / Sonic).

  • Li.Fi bridge integration: cross-chain asset transfers to streamline user onboarding.

  • Modal centralization: unified scattered deposit / withdrawal / account modals into a single AccountModal with a tab-based UI, improving UX.

  • Dual-account system: managed deposits and withdrawals for both Classic (Perpetuals) and Moon (1000x) accounts through one interface.

• Event & competition systems

  • Abstract Cup competition page: integrated a real-time ranking API, team leaderboards, and an end-of-event countdown timer.

  • On-chain upvoting: wagmi-based smart-contract integration with signature verification and duplicate-vote prevention.

  • 3D team carousel: a rotating carousel built with CSS 3D transforms, with team-selection and join interactions.

• AI agent system

  • Multi-agent pipeline: a three-stage Screening → Character → Evaluator structure to keep win/loss judgments consistent.

  • Prompt-injection defenses: tool-calling-based filtering of malicious input, keeping user input separate from the system prompt.

  • 5-level score gauge: color mapping by score range, with a loading animation.

• Performance optimization

  • Eliminated redundant API calls by caching leaderboard and history data and deduplicating repeated authentication requests.

  • Fixed infinite-loop bugs by halting runaway on-chain read requests and preventing duplicate game-creation and bet events.

  • Improved memory management by capping the number of chart data points and merging nearby points to automatically prune stale data.

• Security & incident response

  • Clickjacking protection: blocked iframe embedding via "X-Frame-Options:DENY" and "CSP frame-ancestors".

  • CVE-2025-55182 emergency patch: shipped a Next.js upgrade the same day a React Server Components vulnerability surfaced.

  • BFCache fix: resolved a WebSocket-not-restoring issue on Safari back-navigation by detecting the event.

  • Per-chain contract hotfix: shipped a fix for an incorrect USDC address on withdrawals.

• Code quality

  • Large-scale refactor: split a 2,003-line component into 4 custom hooks and 6 components (down to 273 lines—an 86% reduction).

  • Coordinate utilities: consolidated scattered transformation logic into a single CoordinateSystem class.

  • Custom-hook pattern: extracted business logic into reusable hooks

• Landing pages

  • Interactive teaser page: a 40×40 tile grid built with CSS 3D transforms, with a magnification animation on hover.

  • Service overview page: Hero / About / Metrics / Roadmap sections with scroll animations via IntersectionObserver.

• Third-party payment integrations

  • Developed and integrated partner payment APIs for Sikkwondaejang (a Korean corporate meal-voucher platform), including KakaoPay and Toss Pay.

  • Built the end-to-end flow from kiosk barcode input through external API request, response-code parsing, and persisting payment data to the database.

  • Coordinated directly with the external vendor to align on API specs and drive the integration forward.

• In-house ERP development

  • Developed a product-disposal API for unmanned store operations.

  • Designed the database tables for managing disposal records.

  • Implemented inventory syncing and disposal-reason logging as part of disposal processing.

• System stability & developer experience

  • Introduced Sentry-based error monitoring, replacing a one-second log-scanning routine with real-time error detection.

  • Built real-time error alerting for the team through a Slack integration.

  • Resolved MySQL deadlocks by removing database triggers and refactoring the logic into application-level SQL.

  • Documented the API with Swagger / OpenAPI.

  • Wrote test coverage with Jest.

• Built an end-to-end pipeline from data collection to deployment

  • Implemented a pipeline spanning new-product data collection, preprocessing, image-classification model training, and automated model deployment.

  • Developed a controller that lets operators trigger and monitor the data-collection process through barcode input.

  • Built a training module that governs deep-learning model training and the rollout of updated models.

  • Developed an updater that handles the communication required to deploy revised models.

• Resolved vision-model performance degradation using CycleGAN

  • Proposed a method to close the domain gap between training and test data caused by differences in capture equipment and by data synthesis.

  • Achieved 0.2% higher accuracy than baseline using CycleGAN-rendered data.

  • Combined augmented data with CycleGAN-rendered data to improve accuracy by 8.38% over baseline.

  • Published and presented a paper at KCC 2023 and filed an internal patent application.

• Improved vision-model performance through data augmentation

  • Proposed an approach for using augmented image data to train vision models (detection and classification).

  • Reduced overfitting in the existing deep-learning model through data augmentation, improving accuracy by 22.54%.

  • Built a data collector that cut image-collection time to 30 seconds per SKU, easing data-collection cost.

• Completed coursework in frontend, backend, data analysis, and AI; built 1 individual project and 2 team projects.

• Double Major Earth System Sciences and Physics

• GPA 3.58/4.5

• Extracted and organized band-gap data for SnS (tin sulfide) using angle-resolved photoemission spectroscopy (ARPES) at the Pohang Accelerator Laboratory.

• Presented research papers and led lab study sessions.

• Conducted numerical modeling research with COMSOL Multiphysics.

• Authored a paper on how parameter-dependent groundwater circulation affects geothermal activity at the Ulleungdo volcano.

• Withdrew

A card game platform for predicting Yapper rankings, built to address spam and low-quality content within the Kaito-based InfoFi ecosystem. Users purchase cards to predict influencer rankings, generating high-quality data that feeds back into Kaito.

• Web3 authentication

  • Implemented Twitter OAuth with automatic embedded-wallet creation via Privy.

  • Integrated dual-token backend authentication (access token + identity token).

  • Real-time wallet-balance lookups on the Base Sepolia chain (viem RPC).

• Card-pack opening system

  • Integrated ERC-721 smart-contract minting (ethers.js).

  • Three-stage card-opening animation: Draw → Flip → Reveal.

  • 150-particle confetti effect using Poisson disc sampling.

• Tournament prediction system

  • Tournament registration by league (1–4) and a five-player Yapper team draft.

  • Pick-rate-based Yapper selection UI with multi-region filtering (Total / Korean / Chinese).

• Performance optimization

  • Prevented redundant leaderboard / pick-rate API calls with 1-minute TTL in-memory caching.

  • Map-based cache management per league × region combination.

A service that infers the most similar plant from a catalog of 98 species based on a user-uploaded image.

• Team Lead — organized project meeting documentation and delivered presentations.

• Built an image dataset of 98 plant species through web crawling.

• Pre-trained on PlantNet data and fine-tuned the model to mitigate overfitting and improve performance.

• Selected the classification network by comparing confusion matrices and F1 scores across models including ResNet, Inception-v2, and Xception.

• Grand Prize winner.

  (Project page) (Github)

A project to plan a service themed around COVID-19 and OTT (Github)

• Planned the web service.

• Built a film-festival dataset from open APIs.

• Developed recommendation logic based on festival prestige and award history.

• Selected as an Outstanding Project.