Worked on stabilizing the Gemma3VL training pipeline within the NVIDIA/NeMo repository, addressing training interruptions by refining model export processes and ensuring backward compatibility for dataset batch loading. Applied deep learning and model training expertise to correct vocabulary size handling, preventing export-time mismatches and improving workflow reliability. In the GoogleCloudPlatform/vertex-ai-samples repository, authored a user-facing blog post on mitigating catastrophic forgetting in Vertex AI Training Cluster, focusing on dataset mixing strategies, experimental setups, and evaluation metrics. Leveraged Python and Markdown to enhance documentation, share actionable guidance, and align content with AI best practices, supporting practitioners in optimizing machine learning workflows.