I'm Josh Euikwang Josh

Post-trained GPT-4.1-nano for ternary fact verification using supervised fine-tuning and Direct Preference Optimization, raising macro accuracy from 0.642 to 0.928 across 20 evaluated configurations. Built an RLAIF pipeline generating 14K teacher-distilled training examples and three rounds of DPO preference data from GPT-4.1-mini, then ran ablations across three justification formats and 14 fine-tuning configurations, lifting recall on the weakest class from 0.14 to 0.97.

Implemented PPO from scratch in a fork of the rsl_rl library and designed the reward, observation space, and reset curriculum for a 7-gate racing track, training 8,192 parallel environments in NVIDIA Isaac Lab to 4.0 laps per episode. Deployed the trained policy to a physical Crazyflie quadrotor over a ROS2 and Vicon motion-capture pipeline, using domain randomization to transfer sim-trained control to held-out real-world dynamics.

Built a multithreaded terrain streaming pipeline in C++17 and OpenGL, splitting chunk generation into block-fill, mesh-build, and GPU-upload stages across a QThreadPool to eliminate frame hitches during on-demand streaming. Diagnosed and fixed a shutdown-time use-after-free across worker threads using a shutdown flag, thread-pool drain, and mutex-guarded swap-and-drain, resolving intermittent crashes the team had hit for two weeks.