LLM Agents Help Win Kaggle Competition with 600K Lines of Code

In March 2026, a team leveraging generative AI agents secured first place in a Kaggle Playground competition on telecom customer churn prediction. Using three large language models (LLMs)—GPT-5.4 Pro, Gemini 3.1 Pro, and Claude Opus 4.6—the team generated over 600,000 lines of code, conducted 850 experiments, and built a four-level stack of 150 machine learning models to deliver the winning solution.

The key to their success lay in accelerating the data science workflow. Traditionally, two bottlenecks slow down machine learning experimentation: writing code for new ideas and executing those experiments. GPU acceleration has addressed execution speed, but generative AI agents are now solving the code-generation problem, allowing for rapid prototyping and iteration. This combination is proving to be a game-changer in competitive data science.

How Generative AI Agents Transformed the Process

The Kaggle competition tasked participants with predicting customer churn, with performance assessed by AUC (area under the curve). The winning team followed a robust, structured workflow guided by LLM tools:

• Exploratory Data Analysis (EDA): LLMs analyzed the datasets to identify key features, missing values, and target variables. This step included writing and executing Python scripts iteratively to refine insights.
• Baseline Model Development: LLMs generated code for initial models using algorithms like XGBoost and neural networks. These models provided a starting point for further refinement.
• Feature Engineering: The agents tested various transformations and optimizations to extract stronger signals from the data, continuously iterating on what worked.
• Model Stacking: Experiment results were aggregated into a multi-layer ensemble, combining the strengths of diverse models to maximize prediction accuracy.

By automating repetitive tasks like code generation and testing, the team could focus on strategic decisions and creative problem-solving. The result was a highly performant model stack built in record time.

LLMs in Data Science: A Growing Trend

Large language models are increasingly being integrated into data science workflows, as seen with tools like AutoKaggle, which uses multi-agent AI systems to tackle complex competition tasks. These systems excel at automating data cleaning, feature engineering, and even reading academic papers or forums to generate new ideas. According to recent insights, this shift is not limited to competitions but also extends to broader software development, where LLMs are automating debugging, test case generation, and code optimization.

However, challenges remain. Issues such as code hallucinations, misinterpretation of task context, and security vulnerabilities in AI-generated code require human oversight. Despite these limitations, the rapid adoption of LLMs signals their potential to reshape industries reliant on data-driven decision-making.

Implications for Developers and Data Scientists

The Kaggle success story demonstrates how generative AI can dramatically enhance productivity in data science. For developers, this means a shift from manual coding to high-level tasks like designing workflows, interpreting results, and managing AI agents. NVIDIA’s GPU libraries, such as cuDF and cuML, further accelerate this process, enabling faster execution of AI-generated experiments.

For those looking to replicate or adapt these methods, NVIDIA provides extensive resources, including the CUDA-X for data science libraries and workshops on feature engineering. As the tools and techniques evolve, staying ahead will require leveraging both generative AI and robust computational frameworks.

The bottom line? The data science arms race is being defined by speed and scalability, and LLM agents are rewriting the rulebook for what’s possible in machine learning competitions and beyond.

Image source: Shutterstock