Applied NLP / AI researcher building grounded, evidence-driven language systems. My work connects retrieval and evaluation research with production constraints, informed by earlier research on how brains and language models represent meaning.
Portfolio site · Case studies · Hugging Face · ORCID · Email
Designed a verified-unit QA system that returns an operator-approved answer or abstains. I built the retrieval and evaluation harness, then replaced an inflated result after finding question-identity leakage in the original setup. The current held-out evaluation reports hybrid retrieval with query augmentation at approximately R@1 0.881–0.930.
Fine-tuned and evaluated two Korean retrieval encoders on an H100 workflow. Both model cards and benchmark artifacts are public; on the documented AutoRAG evaluation, the BGE-M3 variant reports MRR 0.7773 and Hit@10 0.9474.
BGE-M3 model · Qwen3 model · Case study
Built an end-to-end on-device rPPG prototype: training pipeline, Core ML export, and SwiftUI application. With a strict participant split, fine-tuning reduced the documented UBFC heart-rate MAE from 5.63 to 2.80 bpm. This is a wellness research prototype, not a medical device.
Code and reproducibility notes · Model
Before my current applied NLP work, I conducted research on metaphor, reasoning, and representational alignment between human neural data and language models. I have research experience at Seoul National University and Korea University; these were research roles, not degree programs.
- Metaphor in Mind and Machine, OHBM 2024 / Aperture Neuro
- Comparative Analysis of Brain and NLP Models for Reasoning Tasks, BESK 2023
- Korean Twitter Bot Detection based on Deep Learning, KSC 2022
Research details and authorship
pycag is an alpha Cache-Augmented Generation toolkit for pre-filling Llama-family models with a bounded knowledge corpus and reusing the resulting KV cache across queries. It is available on PyPI.
Python · PyTorch · Hugging Face · sentence-transformers · hybrid retrieval · RAG evaluation · vLLM · LoRA/ORPO · Docker · Postgres/pgvector · SLURM/H100
I care about grounding, faithful evaluation, reproducibility, and knowing when a model should abstain.