Winning a Kaggle Competition with Generative AI–Assisted Coding

A first-place Kaggle playground competition win was achieved using three LLM agents (GPT-5.4 Pro, Gemini 3.1 Pro, Claude Opus 4.6) that generated over 600,000 lines of code and ran 850 experiments for telecom churn prediction. The workflow follows four steps: EDA, baseline model building, feature engineering, and model combination via hill climbing and stacking. GPU-accelerated libraries (cuDF, cuML, XGBoost, PyTorch) handled fast execution while LLM agents handled fast code generation. The final solution is a four-level stack of 150 models. Specific prompts are provided for each workflow step, showing how to guide LLM agents through the full ML pipeline.