How Bayanii turns a spreadsheet into a model

Bayanii works with three kinds of problems, and routes your data based on the column you choose to predict and its type:

• Regression — predicting a number, like price or demand.
• Classification — predicting a category, like churn yes/no.
• Forecasting — predicting future values of something measured over time.

Forecasting is treated as its own kind of problem, not regression with a date column. The rows are kept in time order and never shuffled, and the model predicts forward from past values — because in a time series, when something happened is part of the signal. Regression and classification, by contrast, treat each row as independent.

Before any model is trained, the dataset goes through around thirty automatic checks. The idea is simple: a model that scores well by copying the answer is worse than one that admits the data is not ready. Each issue lands in one of three buckets.

A couple of details worth naming: an ID-like column is caught by being unique on every row and steadily increasing, and Arabic-Indic digits are recognised and converted when Bayanii reads numbers and dates.

There is one rule the whole preparation step follows: every transformation is learned from the training data only, then applied to the held-out and future data. This keeps the score honest, because the model never gets a peek at the data it will be graded on.

• Rows with a missing target are dropped before the split; remaining gaps are filled in (the middle value for numbers, the most common value for categories).
• Column types are cleaned up — a column counts as numeric when at least seventy percent of its values parse as numbers (currency symbols and stray text are stripped, Arabic-Indic digits converted); otherwise it is treated as a category.
• Categories are encoded: one-hot for up to ten distinct values, frequency encoding above that. For classification, the target is label-encoded.
• When there are more than ten features, an adaptive step selects the most useful ones — the method depends on how many features and rows you have — and is fit on training data only.
• Regression and classification use an 80/20 split with a fixed seed so results are reproducible; forecasting holds out the most recent rows, in time order.

For each task, seven model families are trained in parallel rather than asking you to pick one:

Each model is scored by cross-validation — repeated five-fold for regression and classification, so the score is an average across many splits rather than one lucky one. Forecasting uses time-ordered backtesting that respects chronology, then refits the chosen model on the full history. For forecasting, each model family is wrapped in a recursive forecaster that predicts one step at a time using its own recent predictions.

The metrics depend on the task: regression reports R², RMSE and MAE; classification reports accuracy, precision, recall, F1 and a confusion matrix; forecasting reports MAE, RMSE and MAPE. Imbalanced categories are handled with balanced class weights rather than throwing data away.

Every model gets a single composite score between 0 and 1 (higher is better) that blends the task's metrics into one comparable number — so a balanced model is rewarded over one that is strong on a single metric and weak elsewhere.

Models that finish within a small margin of the leader — effectively a tie — get a budget-aware hyperparameter search. The search size is set by how much time is left before the job's deadline, and it is skipped gracefully if there is no time, no settings to tune, or it errors out. The tuned winner is then re-scored the same way as the rest.

Finally, the winning model's metrics are recomputed on a held-out test set, on the original scale of your data — so the numbers you read are in your units, not scaled ones.

This is auditable rather than a claim: every trial is logged — which algorithm, whether it was tuned, its settings, its cross-validation scores, its composite score and timing — and that record is what fills the model card's "Model selection" and "Hyperparameters" tabs.

Bayanii saves the whole pipeline, not just a model. Alongside the chosen model (or, for forecasting, the forecaster and its time index) it stores every fitted step: the scaler, the encoders, the target encoder, the feature selector and the list of selected features, the polynomial step, and the imputer — plus the training-validation record behind the charts on the model card.

Why keep all of it? At prediction time, your new data has to travel through the identical steps in the identical order it did during training. A saved model on its own would not know how the data was scaled or encoded, and the results would drift.

The schema and metadata — column names, target, task type, metrics and the chosen algorithm — are versioned. Retraining on the same name creates a new version while the previous one stays available.

When Bayanii writes a plain-language explanation of your dataset, it sends only summary statistics, column names, and the most common category labels to the AI service — never your raw rows. The underlying records stay put.

Training and prediction run on Bayanii's own infrastructure, separate from the explanation model. See the Privacy page.

• Web app — Flask, served behind a production web server and reverse proxy, with Jinja templates and Tailwind CSS. Sign-in is handled by Firebase Auth, with CSRF protection and rate limiting on top.
• Training — runs out of band on an asynchronous job queue (Redis-backed) with a separate pool of worker processes, so the web app stays responsive while models train.
• Machine learning — pandas, NumPy, scikit-learn, XGBoost, CatBoost, skforecast and joblib.
• Storage — trained pipelines in object storage, metadata in a managed SQL database, and the job queue and live progress in Redis.
• Hosting — a multi-tier private cloud network on AWS, based in the EU.

• Bayanii is in beta — please don't upload sensitive or personal data.
• It works with tabular CSV data only — not images, free text, or audio.
• It will not invent signal that is not in your data. A clean, honest process can still conclude that your columns do not predict the target — and that is a real answer, not a failure.
• A model finds patterns and associations, not causes. A strong prediction is not proof that one thing causes another.
• Predictions are only as good as the data behind them. Quality and accuracy depend on how representative and clean your dataset is.
• Forecasting needs at least one hundred observations and supports horizons up to twenty-four steps.
• Defaults favor a fast, honest baseline over exhaustive tuning.

Have a question this page did not answer? Read the FAQ or contact us.