Work
Built two computer vision modules for insurance claim data, each delivered as an end-to-end pipeline with a web service.
Developed the backend of a social networking service with NestJS, and built the crawlers and analysis that backed product decisions. Completed the 2025 I-Corps program, including three weeks of entrepreneurship training at UC Berkeley.
Building and operating the institute's research platform from
scratch. Eight services now run at *.knpu.re.kr.
In April 2024 the LLM-based legal complaint drafting service
earned a Letter of Appreciation from the university.
Projects
Searches DBLP publication data for faculty candidates using Grok-based name scores. Includes researcher profiles, rankings, monthly updates, and email subscriptions.
Automatically assigns consultation times from participants’ availability using min-cost max-flow. Supports guest responses, Google Calendar, and encrypted consultation records.
Automatically masks every vehicle license plate in dashcam footage. YOLO vehicle detection feeds a degradation-tuned plate detector, sped up with optical-flow tracking and parallel segment processing.
Detects document regions edited with generative AI. Built FinForge, a 6,000-image forgery dataset of insurance claim forms, and fine-tuned TruFor on it to localize edits and score each document.
A dataset browser for original and AI-forged insurance claim documents. Compare image pairs, inspect edited regions and generation metadata, and explore dataset statistics by generation model.
Desktop research workspace for browsing datasets, running collection and analysis jobs, and managing users. Extended its workflows to web services with shared job logs and LLM settings.
Collects research data from news, blogs, cafes, and YouTube with asynchronous I/O, proxy rotation, and priority queues. Supports progress tracking, cancellation, and resume.
Builds and visualizes relationships between words and entities. AI explains graph metrics, keywords, and communities with links to supporting source text.
Runs statistical analyses and visualizes tables and charts. Adds evidence-based AI reports and a Q&A agent that searches tables, reads results, and calls calculation tools.
Implements KEM/KIM keyword signal analysis with adjustable periods and weights. AI interprets signal changes using related terms and representative source articles.
Transcribes audio through a GPU server and manages results and job progress. Limits concurrent transcriptions and retries transient connection failures.
Turns an incident description into a structured criminal complaint draft. Supports guided input, Word/PDF generation, previews, and downloads.
Answers public questions about the lab and guides researchers through its services. A tool-using agent searches documents and code, reads sources, and explains workflows with evidence.
Connects 10 public datasets to help sports voucher recipients find suitable classes and administrators review regional supply gaps and facility candidates. A tool-using AI agent retrieves data and explains calculated results with evidence.
Connects lecture slides, recordings, and board notes with AI explanations. A chatbot for each slide uses its image, explanation, and lecture outline to answer follow-up questions.
Monitor and control Claude Code, Codex, and development tools from a PC or phone. I use UnivDash for all my development work, including agent conversations, files, Git, and server monitoring.
Deploy local LLMs through a CLI and web console. Automates model downloads and Docker/vLLM startup, estimates GPU memory including KV cache, and monitors jobs and reservations.
AI phone-call English practice. Call an AI conversation partner and talk in real time — interrupt, pause, hang up — then get a speaking evaluation report after every call.
Aggregates POSTECH internship and recruitment postings with search filters, applicant-count history, and a deadline calendar. Automates SES collection and SSO re-login with Playwright, stores updates in MongoDB, and emails new postings and watchlist changes.