Youngjae Cho

alternative military service

• Own the LLM/VLM auto-labeling system for content-safety scoring (5 categories × severity): mined the implicit decision tree from 299K VLM reasoning traces and turned it into auditable, tree-structured prompts — macro F1 0.777 → 0.857 (precision +0.103, recall +0.050) and over 4× the labeling throughput of the process it replaced, retiring logic that lived unwritten in each labeler's head
• First-author alignment research (GAPO) — geometric anchoring that stabilizes preference optimization under noisy labels, +3.6pp AlpacaEval 2.0 LC over SimPO
• Winner — NVIDIA Nemotron Hackathon, Track B (Domain-Specialized Model), 2026: built the RL post-training framework for Nemotron-Nano-12B-VL, an architecture no existing RL framework supported — added the RL layer (rollout, loss, training loop) onto Megatron-Bridge with SimPO as the objective, and open-sourced it
• Redesigned the serving layer of the production content-moderation stack so a model swap is a config change rather than a serving-code rewrite — shipped a new inference task type end to end on Ray Serve (schema, strategy, batching, HTTP ingress, tracing), moved calibrated decision thresholds into the checkpoint artifact, and made deployed feature flags verifiable from telemetry
• Built the LLM benchmarking CLI and pluggable engine abstraction (vLLM / Hugging Face / synthetic CI fallback) the team evaluates on

alternative military service

• Diffusion-based defect synthesis for industrial anomaly detection — background-aware disentanglement; best FID/LPIPS over DFMGAN and AnomalyDiffusion on MVTec-AD and LOCO, and higher detector precision/recall on the production line
• Shipped real-time OCR for steel-plate IDs on NVIDIA Triton — ONNX→TensorRT, CUDA graphs, and test-time augmentation for outdoor CCTV conditions
• Made diffusion finetuning and inference fit a single-GPU budget with Flash-Attention + DeepSpeed (FP16)

advised by Il-Chul Moon

First-author work at ICML 2023 (SAAL) and AAAI 2024 (APP) — active learning under scarce labels, and Bayesian prompt adaptation under distribution shift.

Domain-Specialized Model track. Implemented the preference-optimization post-training (SimPO on Megatron-Bridge) behind the team's winning video content-safety VLM.