Drop a paper.
Get working code.
Four ADK agents analyze the paper, plan a minimal implementation, write complete PyTorch code, and repair it — producing a runnable Jupyter notebook in minutes.
Two nn.Module implementations: the competing method and the paper's proposed approach — trained side by side.
Title, intuition, setup, data pipeline, models, training loop, experiment, inference, visualizations, and reflection.
A dedicated review agent catches undefined variables, missing imports, and shape mismatches before you run a single cell.
Architectures are reduced to train in minutes on a laptop — preserve the core idea, skip the GPU cluster.
Download the generated draft while validation runs in the background. Both versions available.
One click publishes the notebook as a GitHub Gist and opens it directly in Google Colab.