MLStudio

MLStudio lets people train, validate, and reuse regression and time-series models through a guided interface instead of model-specific code. Its original Tkinter workbench grew out of my consulting work and was used across 30+ organizations, cutting time to a first validated model from 10 working days to 2. I later rewrote it in Streamlit/PyTorch, separating the modeling backend from the interface and adding time-aware validation and versioned model bundles. Both figures refer to the desktop workbench.

What it does

• Supervised models: train, validate with random, ordered, or cross-validation splits (optionally with grid search and feature selection), or test a saved model on new data. Regressors include Random Forest, Gradient Boosting, SVR, XGBoost, CatBoost, and voting ensembles.
• Neural time series: MLP, CNN, RNN, GRU, LSTM, Bi-LSTM, and ConvLSTM models over lag windows chosen by index or by autocorrelation.
• Metrics: R², MAE, RMSE, and MAPE wherever labels are available.
• Model bundles: export a fitted model with its preprocessing and feature requirements, then reuse it on new data.

Why the rewrite

In the desktop version, interface and modeling code were one class, so every client-specific change risked the models. The rewrite’s main change is a boundary: the modeling backend has no Streamlit dependency, and an architecture test fails if one appears. The interface can change without moving model logic into page code, and bundles carry preprocessing, feature requirements, and model state, so prediction runs exactly the steps training did.

Evaluation and limits

Ordered data uses time-aware splits and folds, and recursive forecasts feed predictions back into the model instead of peeking at actual future values. MAPE skips zero targets, where it is undefined.

Missing-value imputation is not implemented yet. Bundles use Python serialization, so load only trusted files built with compatible library versions.