Hamin Jeoung

I study natural language processing in the Department of Software at Yonsei University Mirae Campus, where I am a fourth-year student. With three co-authors, I built a benchmark that evaluates language models on the Korean Medicine licensing examination, and we are revising the paper. Since September 2025 I have been a research assistant on a National Research Foundation of Korea project on AI for early prediction of aging-related diseases.

Most of this site is written in Korean. 한국어

Projects

  • Evaluating Language Models on the Korean Medicine Licensing Examination

    As a team of four authors, we built a benchmark of 1,138 questions from four years of the Korean Medicine licensing examination, along with retrieval resources. Nine open models averaged 36.00% without retrieval, and none of the open models we evaluated reached the 60% overall score required to pass.

  • Nongjian: Satellite and Weather Data for Crop Disaster Loss Assessment

    A platform for agricultural disaster insurance that reads crop damage from satellite imagery and weather data before field inspections and produces assessment documents. The team has a working prototype from parcel lookup to report output. It won the top prize for idea planning at the 2026 Yeongcheon public data and AI startup competition and was selected for the Ministry of Education's Student Startup 300+ program.

  • People Who Buy Cars Because There Is No Bus

    Using public data for Korea's 228 districts and 1,173 rural townships, our three-person team analyzed forced car dependence where buses are scarce and where 2026 national transport subsidies go. In rural counties, 97% of the per-capita benefit from the two national programs came from the fuel tax cut, and our models suggest that better public transit alone would barely reduce car ownership there.

  • Beddy 2.0: Safety Layers for an AI Companion for Hospitalized Children

    Beddy is an AI companion app for children in long-term hospital care, first built by another team at a hackathon. Our v2 replaces its crisis check with a five-level one and adds other child-safety layers, including a leak check on text sent to guardians and guardian consent for children under 14. On synthetic test sentences, the guardian false-alarm rate on everyday messages fell from 8.2% to 3.5%, and leak detection rose from 0% to 72.0%.

  • Korean Sign Language Word Recognition with MediaPipe and DTW

    A webcam demo that turns MediaPipe hand and face landmarks into a 102-dimensional feature per frame and picks a Korean Sign Language word by DTW distance to per-word prototypes. Each of the five words (bear, construction site, park, factory, playground) has prototypes from a single AI Hub sign video. Park and playground ended up with identical landmarks, and there is no held-out evaluation.

  • Porting the TS-CAN rPPG Model to NVIDIA DeepStream

    Hackathon code with a script that converts TS-CAN, a public rPPG model that estimates the blood volume pulse from face video, to ONNX, and an NVIDIA DeepStream app meant to feed the model the 10 most recent crops of detected people. The model code, training script, trained weights and sample data come from the public repository ubicomplab/deep-rppg-android.

Contact