I work on geometric deep learning, spatio-temporal modeling, and AI for Science. I build models that see the world as graphs and sequences, then ship them as real production systems.
I build AI that can make sense of complex, real world systems and predict how they behave.
I hold an M.S. in AI from Ewha Womans University, where I focused on geometric deep learning, spatio-temporal modeling, and AI for Science.
My research covers graph neural networks and physics-informed learning, and I apply them to large physical and human systems, from biomechanical deformation all the way to nation scale EV charging demand.
Beyond research, I've shipped production AI at three startups, most recently as the founding AI engineer at VStdy ↗.
▸ spatio-temporal GNNs · geometric deep learning · physics-informed learning
▸ LLM agents & adaptive learning · IRT · multi-agent orchestration · RAG · fine-tuning
▸ AI for science · biomechanics · human motion · large-scale time-series
Built an 11-model STGNN benchmark behind one training and evaluation interface: 33 jobs across energy, occupancy, and arrival targets, 168h → 48h forecast.
State-machine orchestrator coordinating six specialist agents; per-topic mastery score and a Scan&Solve pipeline (Mathpix → GPT-4o → TTS). Sole engineer on the SwiftUI iOS app, FastAPI backend, and Next.js site.
Treated the 3D foot mesh as a graph and learned a per-node displacement field that works as a fast surrogate for the FEA solver, with continuum-mechanics regularizers and subject-adaptive conditioning (M.S. thesis).
Proposed the triplet-loss encoding approach and pipeline and wrote the experiments section: FFT of 2D gait keypoints fused with IMU data.
Fine-tuned LLaMA 3.1 (8B) on a syllabic-template + context format, optimized with three simultaneous objectives: syllable, context (BERT), and rhyme-score losses.
Reframed cat ID as a metric learning question, basically "are these two photos the same cat?", using triplet-loss contrastive backbones (ResNet, MobileNet, ViT) benchmarked for IoT.
More research and engineering projects. Full code, papers, and demos available on request.