An open-source tool to create, train and use neural networks in no-code (github.com)

MarijoAI is an open-source, browser-based no-code tool that lets you build, train and run simple neural networks entirely on your device. The app provides a visual, step-by-step workflow (Clean Data → Create → Train → Predict) for CSV parsing/validation, architecture definition, in-browser training with a custom JavaScript neural-net implementation, and local export/import of trained models as JSON. All processing is client-side (no server uploads); you run it by opening index.html from the repo and your data never leaves your machine. The project is MIT-licensed and includes example data and exercises for learning. Technically, MarijoAI implements common activation functions (ReLU, Sigmoid, Tanh, Softmax), the Adam optimizer, early stopping, and live training history (loss/accuracy). It uses PapaParse for CSV handling and bundles the Wisconsin Breast Cancer dataset for tutorials. After training you can evaluate accuracy when ground truth is available, export predictions, and reload saved models for subsequent offline inference. The tool is most useful for education, hands-on experimentation, and privacy-sensitive prototyping where lightweight, CPU-bound models suffice — a practical sandbox for learning core ML concepts without installing frameworks or sending data to the cloud.