I build learning systems that don't fall apart when supervision gets messy.
Research on alignment, robustness, and optimization geometry — and the pipelines that ship it. From proofs to TensorRT serving.
ML Research Scientist @ Pyler · ex-Aiv · M.S./B.S. KAIST, advised by Il-Chul Moon
A throughline runs through the work: local geometry — sharpness, curvature, anchors — as a practical proxy for when a model can be trusted.
All publications →Replaces DPO's frozen reference with a dynamic adversarial anchor — a worst-case local perturbation of the current policy — and reweights each preference pair by its geometric brittleness. Robust to noisy labels without an explicit noise model.
Bayesian, data-dependent priors that let vision-language prompts adapt per input instead of staying fixed — better calibration under distribution shift.
Read the paper →An acquisition function that bridges sharpness-aware minimization and active learning — selecting samples that flatten the loss landscape.
Read the paper →Diffusion synthesis that disentangles defect from background via masked cross-attention — the method behind the production system on the projects page.
Read the paper →Across Pyler and Aiv — the method, the training and inference efficiency, the serving stack, and the proof each one moved the metric.
Full case studies →Implemented reference-free RLHF (SimPO) for a 12B video-language model the framework couldn't post-train — team won Track B.
Case study →Reverse-engineered the implicit labeling decision tree from 299K VLM reasoning traces.
Case study →Diffusion defect generator that disentangles defect from background; best FID/LPIPS vs baselines.
Case study →Two-stage Triton OCR pipeline, optimized to real-time with TensorRT + CUDA graphs.
Case study →Winner — NVIDIA Nemotron Hackathon, Track B (Domain-Specialized Model), 2026
Diffusion-based defect synthesis pipeline for industrial anomaly detection (background-aware disentanglement)