Projects
I treat real-world applications as sources of research questions. Across industrial QA, multimodal information access, foundation models, and long-form video, I look for technical bottlenecks that prevent AI capabilities from translating into practical value.
Multimodal Reasoning & Document AI 5
Video Curation System
Video QA and RAG system for structured curation and grounded retrieval over long multimodal content.
DocGraph Copilot
Structure-aware multimodal chunking pipeline for grounded retrieval and long-document QA.
Science-Domain Korean LLM & RAG Pipeline
Korean LLM and RAG pipeline specialized for the science domain; led data preprocessing and chunking methodology development, producing an EMNLP 2024 Industry Track paper and 2 technology transfers.
Synerpeace Pilot
Insurance document QA pilot combining LLM-based retrieval with production-oriented application development.
- Need
- Reduce the time staff spend searching complex insurance documents.
- Research problem
- Conditions, exceptions, and provenance were lost during retrieval.
- Outcome
- ~40% reduction in manual search time; motivated subsequent structure-aware retrieval research.
Synapse
Multilingual LLM-based medical consultation chatbot; selected for the KU-Grant research commercialization program.
Foundation Models and Adaptation 6
In parallel with my research on multimodal evidence systems, I contributed to Korean and multilingual foundation models across academic, industrial, and consortium settings, including the Adapter-based architecture introduced in Exobrain WiseQA that grounded my later work on structured evidence.
National Representative K-AI Research Team
Multilingual foundation model development as a national-level representative researcher, spanning data pipelines, post-training, and evaluation.
Ko-Gemma
Korean adaptation of Google's Gemma model; contributed Korean-focused data adaptation and Korean benchmark evaluation.
KULLM
Korean LLM initiative developed at Korea University; contributed instruction-following tuning and benchmark evaluation.
Robot Process Automation – PLC Control-Code Generation
GPT-based assistant that generates PLC control code for smart-factory robot process automation, in collaboration with Hyundai Mobis.
Mi:deum K 1.0
Contributed to KT's Korean LLM ecosystem for enterprise generative AI; led model adaptation and Korean-targeted post-training evaluation.
Exobrain WiseQA (ETRI)
Table QA and neural-symbolic Korean sentence analysis for the national Exobrain project; an Adapter-based architecture unifying text/table embedding models reduced QA model parameter size by about 50%.
Additional Projects 5
Earlier applied AI collaborations spanning Korean speech, education, and geoscience.
TTS Prosodic Boundary Prediction (KT)
Sentence analysis and prosodic boundary prediction for Korean TTS quality improvement (92.13%).
TTS Pronunciation Sequence (KT)
Korean phoneme sequence generation for higher-quality TTS systems, deployed in KT GiGA Genie.
Navigation TTS System (Navice)
Korean phoneme-based navigation TTS system developed with Navice, deployed in BMW and Mercedes-Benz vehicles.
AI Online Judge for Coding Education
AI-based online judge system deployed in the PNU SW curriculum with NextStage.
AI Mineral Location Prediction (KIGAM)
Geoscience AI system for mineral location prediction and domain-specific feature engineering.