Starting and iterating on a Kaggle competition in Google Antigravity (andlukyane.com)

A recent exploration of Google Antigravity 2.0 in a Kaggle competition demonstrated the capabilities and limitations of this AI-driven development platform. The competition focused on predicting health conditions from tabular data, leveraging Gemini 3.1 Pro, a model designed for complex reasoning. The author started the project using knowledge from previous competitions and integrated it into Antigravity, which autonomously planned tasks, executed code, and generated reviewable outputs. However, the experience highlighted critical aspects of effective modeling, particularly the importance of correct cross-validation and evaluation metrics. Despite achieving a high cross-validation accuracy, the initial leaderboard submission underperformed, revealing a discrepancy attributed to incorrect assumptions about the evaluation metrics and label mismatches. After addressing these issues, the author improved the leaderboard score significantly. This exercise illustrated not only the potential for AI-powered tools to streamline workflows but also emphasized the necessity for human oversight and understanding in data science projects. The study serves as a reminder that while AI can facilitate iterative processes, foundational best practices in model evaluation remain vital for success in the AI/ML community.