ML Pipeline

Upload a CSV and get a trained model without writing any code. Four algorithms — Random Forest, XGBoost, LightGBM and CatBoost — compete on 5-fold cross-validation, and the winner is chosen automatically on F1 for classification or MAE for regression. Optional Optuna tuning and a SHAP explanation then run on whichever model won.

Check whether live data has drifted away from what your model was trained on. Upload a new production batch and compare it against the training baseline: PSI, KS test and distribution histograms for numeric columns, category frequency shifts for categoricals. A trend sparkline tracks the drift score across successive batches.

Clean a messy CSV before you train on it. Deduplicate rows, fill missing values with 8 numeric strategies (mean, median, KNN, MICE, forward or backward fill, a constant, or drop the row) or 5 categorical ones, strip outliers with the 1.5 × IQR rule, and correct skew with a log transform. It all runs in your browser. Download the cleaned file, or send it straight through to AutoML.

Combine the strongest models instead of betting on one. Voting (VotingClassifier / VotingRegressor) or stacking with a meta-learner on top of the AutoML winners, which typically reduces variance and generalises better than any single model on its own.

Upload a CSV and see what you actually have before deciding what to do with it. Overview and duplicate count, a per-column profile with missing values and cardinality, distributions and a correlation matrix, principal components and mutual information, and a pass, warn or fail verdict on whether each column is fit to model on. A statistic that is not defined for your data reads as a dash, never as zero. Nothing is stored, and the file is analysed in memory.

Build new features out of your columns without writing code. Per-column transforms (log1p, sqrt, z-score, min-max, percentile rank, winsorising, outlier and missing flags) plus binning, polynomial and interaction terms, ratios, lags and rolling windows, date extraction and cyclical sin/cos encoding. It all runs in your browser, and the CSV it writes uses whole-file statistics — the transformer that ships inside a trained pipeline is the one that fits on training data only.

Cut a dataset down to the columns that actually carry signal. Four methods — variance threshold, correlation filter (drops anything above 0.9), recursive feature elimination with a Random Forest, and SelectKBest on mutual information — prune redundant columns before training, with a configurable top-K cutoff.

Squeeze more out of the model AutoML picked. A TPE sampler runs up to 30 trials searching for better hyperparameters. It runs after model selection rather than before, so tuning can never inflate the score that won the competition in the first place.

Paste a numeric series or a list of event timestamps and it finds the strongest repeating cycle — the 'every 7 days', 'every 24 hours', 'every 12 steps' pattern hidden in the data. A Fast Fourier Transform, run after detrending and a Hann window, turns the series into a spectrum; the dominant peaks are the cycles, each with a strength above the noise floor. Numeric series report cycle length in samples, timestamps in real time units. Nothing is uploaded — it runs entirely in your browser.

Run the whole pipeline as one sequence instead of tool by tool. A visual canvas chains all seven stages together — preprocessing, feature engineering, feature selection, AutoML, Optuna tuning, SHAP explanation and ensembling — so a labelled CSV goes in one end and a trained, explained model comes out the other.

Watch the seven ML stages play out as an animation rather than reading about them. Illustrated characters carry data through each step of the pipeline in turn. Nothing to upload — it is a walkthrough of how the stages fit together.

Watch traffic to this site arrive as it happens. Page views and tool opens flow from the browser into PostgreSQL through a FastAPI ingestion endpoint, and Supabase Realtime pushes each new row straight to the dashboard — no polling, no refresh button.

See why a model made a particular prediction, not just what it predicted. Every result comes with a SHAP bar chart showing which features pushed it and by how much. Engineered columns are grouped back to the original feature they came from, so you read source influence rather than transform noise.