Projects
UCSD Qualcomm Institute — FinBERT vs LLM Sentiment Classification
2026.01~02A 37-day AI research program at UC San Diego's Qualcomm Institute, one of five students selected by Chosun University; served as deputy student lead. On a five-person team drawn from two universities, wrote and ran the BERT and FinBERT baselines and owned the data pipeline — cleaning and language-filtering 650,000 Yelp reviews, keeping 617,611 through a Gemini quality pass, and labeling them through the OpenAI batch and Gemini APIs. Measured negative transfer: FPB-trained FinBERT classified 90–96% of reviews rated 3 or above as neutral, while zero-shot Gemini classified 98.7% of 1-star reviews as negative and 97.6% of 5-star reviews as positive. Kept working with the advisor after the program ended until the study was published at KCC 2026 (second author of six). A follow-up semi-supervised self-training run drew pseudo-labels from the 124,381 samples where FinBERT and Gemini agreed, lifting Macro-F1 on the financial test set from 0.84–0.85 to 0.87 with no manual annotation — this follow-up is unpublished.
ChosunInsight — Hybrid RAG Assistant
2026A campus assistant that answers questions about Chosun University academic records. Intent analysis routes a question to a structured handler, the vector store, or live domain search; dense (ChromaDB) and sparse (BM25) retrieval are fused with Reciprocal Rank Fusion. Built with a five-person industry-academia team; owned the RAG engine. Also wrote the ingestion and index pipeline and made the assistant decline out-of-scope questions rather than hallucinate an answer.
CTF-Solver — Security Automation MCP Server (⭐53)
2025.09~11Extended an MIT-licensed open-source MCP server (Wh0am123/MCP-Kali-Server, ~10 tools) to 55+ callable Kali Linux security tools, added automated forensics, crypto and cloud workflows and Docker packaging, and designed multi-step workflows. Re-scoped it from automated pentesting to CTF problem-solving to limit misuse. A solo continuation of the White Hat School capstone Not-Yet. Published to the Smithery registry, where it drew roughly 600 tool calls a month while it was live.
Not-Yet — Security Tool-Use MCP Server
2025.07White Hat School Cohort 3 capstone (team project, development lead). A FastMCP server fronting two Flask services — one running Kali penetration-testing tools, one wrapping a search API that I added — so an LLM can choose and call tools rather than follow a fixed script. Built on an open-source Flask-tool-server and MCP-shim pattern. → Continued solo as CTF-Solver.
Deepfake Voice Detection Service — Information Security Competition, Excellence Award (2025.04)
2025Backend for a service that judges whether call audio is synthetic. Redesigned a per-request model-reload architecture into a single load at startup plus an asynchronous task queue, and deployed it on Oracle Cloud ARM64. Excellence Award, Mirae & Software Information Security Competition.
Personal Portfolio & Blog
2026~This site. A portfolio and blog built with Next.js — MDX posts, gallery, admin CMS, dark mode — self-hosted over HTTPS with Docker and Nginx.
iPod Classic ithmb Converter
2026.03A Python GUI tool that restores photos off an iPod Classic. Existing ithmb converters were paid, so I built a free alternative, with Real-ESRGAN AI upscaling.
PALLOW — Location- and MBTI-Based Friend Recommendation
2024A location-based social platform with algorithmic friend recommendation, group meetups, real-time chat and interest communities. Led a five-person team; separated a Java/Spring main server from a cosine-similarity Flask recommendation server and deployed both to Oracle Cloud with Docker.
Awards & Activity
- 2026.07
2026 National AI·SW University Competition (AI track) — 28th of 269 teamsDetails ＋
Predicted an agent's next action across 14 classes. 28th of 269 teams (team HUE, Macro-F1 0.79334) under a 10-minute, 1 GB, offline budget. Traced a persistent confusion among 4 classes to label ambiguity rather than model capacity, found 37.6% of inputs being silently truncated, and traced a failed ensemble back to uncalibrated logits.
- 2026.07
★ Excellence Award — Honam Regional SW University LLM HackathonDetails ＋
Overnight hackathon.
- 2026.06
KCC 2026 paper published — FinBERT vs LLMDetails ＋
“Development of Sentiment Classification Models for U.S. Yelp Reviews: FinBERT vs LLM Performance Comparison”, Korea Computer Congress 2026. Second author of six (corresponding author: Dr. Seokheon Cho, UCSD).
- 2026.04~06
Intermediate AI course (AWS cloud architecture), 160 hoursDetails ＋
Ministry of Science and ICT · NIPA / Honam ICT Innovation Square.
- 2026.01~02
UCSD Qualcomm Institute AI Research Program — completedDetails ＋
Deputy student lead of the 16th cohort. Co-authored “Development of Sentiment Classification Models for U.S. Yelp Reviews: FinBERT vs LLM Performance Comparison”.
- 2026~
AI Pilot Top Gun Challenge 2026 — finalistDetails ＋
Air-combat pilot agent on an Unreal Engine simulation (ongoing).
- 2026★ Encouragement Award — Honam IS Hackathon
- 2025.11
2025 LLM Safety Challenge — participantDetails ＋
Solved every posed problem; eliminated in the preliminary round.
- 2025.10~11
★ Top Excellence Award — 2025 Chosun University SW AI ChallengeDetails ＋
Crack segmentation under a 500M MAC budget. Tuned an EfficientNet-B0 + GroupNorm U-Net decoder to 499.5M MACs; 5th on the raw leaderboard and 2nd of 30 teams after three over-cap entries were disqualified.
- 2025.04
★ Excellence Award — 5th Mirae & Software CompetitionDetails ＋
Deepfake voice detection service (backend).
- 2025.03~09
White Hat School Cohort 3 — completedDetails ＋
7-month national security program (Ministry of Science and ICT · KITRI), four-stage selection.