Skip to main content
That is the whole call. There is no training step, nothing to tune, and no pipeline to maintain.

What Hollerith is

Hollerith is a transformer pretrained to predict directly from tables, trained and served by Monarcha. It reads your labeled table at prediction time rather than learning from it. fit computes no gradients. It hands the model your table, and predict scores new rows against it. The name is Herman Hollerith, whose punch card carried the 1890 census and made the first machine-readable table.

What that gets you

  • A result before you build a pipeline. Point it at a labeled table and predict. You find out whether the signal is there before you spend the week on feature engineering.
  • One endpoint instead of many models. You keep one call where you would otherwise keep a trained model per table, per segment or per customer.
  • Nothing to tune. There is no learning rate, no tree depth, no search space and no retraining schedule.

What it costs

  • Your table is read on every call. Its size is a cost you pay each time rather than once.
  • fit is a real job on a GPU. It takes seconds to minutes and bills the rows you send.
  • Column names carry no meaning. Text and dates are read as categories, so the model sees which rows share a value, not what the value means.
“No training” is a claim about gradients. It is not a claim about cost or latency, and How Hollerith works is honest about the difference.

Where to start