Winning a Kaggle Competition with Generative AI–Assisted Coding

LLM Agents Drive Kaggle Win by Automating 850 Experiments

An NVIDIA data scientist and Kaggle Grandmaster, Chris Deotte, recently secured a first-place finish in a tabular data competition by deploying a team of three large language model agents. The system generated over 600,000 lines of code and ran 850 separate experiments to predict telecom customer churn. This result provides a concrete example of how generative AI can be directed to automate the high-volume, iterative work required for top-tier machine learning performance, moving beyond simple code completion to systematic, scaled experimentation.

A Guided, GPU-Accelerated Workflow

The winning solution was a four-level stack of 150 models, which was developed by guiding agents—including OpenAI's GPT-5.4 Pro, Google's Gemini 3.1 Pro, and Anthropic's Claude Opus 4.6—through a structured process. This human-in-the-loop strategy combined the agents' rapid code generation with the execution speed of NVIDIA hardware and libraries like cuDF and cuML. The workflow was methodical and repeatable:

• Data Analysis: Agents first performed exploratory data analysis (EDA) to understand the dataset's structure, features, and target variable.
• Baseline Modeling: LLMs then generated initial pipelines for a variety of models, such as XGBoost and neural networks, to establish performance benchmarks.
• Feature Engineering: The system iteratively generated and tested new features to improve model performance, saving the outputs of every experiment.
• Ensembling: Finally, agents were tasked with combining the hundreds of resulting models using advanced techniques like stacking and hill climbing to build the final winning solution.

This case study shows how modern machine learning development can overcome its two main bottlenecks: the speed of writing code and the speed of running experiments. By pairing LLM agents with GPU acceleration, data scientists can substantially shorten the cycle from idea to validated result. The approach suggests that a practitioner's effectiveness may soon be measured less by their line-by-line coding speed and more by their ability to strategically manage these automated, high-throughput systems.

The competitive advantage in machine learning is shifting from manual coding prowess to the strategic orchestration of AI agents and accelerated hardware. This workflow doesn't replace the data scientist; it equips them with a system to test hundreds of hypotheses in the time it previously took to test a few, elevating their role to that of a strategist directing an automated research team.