Quick Prediction
You already have the rows. Drop them in.
When your data is already collected and already labeled, you do not need a pipeline. Upload the rows you have judged, upload the rows you have not, and get a scored column back. If the data or the labels are the missing part, use Auto Prediction instead.
How it works
Two files in, one column out
Upload labeled examples
A CSV where one column holds the answer: a label, a grade, a score. Everything else is context the model can use.
Upload rows to score
A second CSV with the same feature columns and no answer column. Shared columns are matched for you; extras are ignored.
Get a predicted column back
Your second file comes back with prediction appended, plus a confidence figure for classification. Preview it, then download the CSV.
The model
A foundation model, not a training run
Traditional machine learning needs a training run for every new question: pick an algorithm, fit it, tune it, validate it, then do it again when the question changes. That is weeks of work and a person who knows how to do it.
A tabular foundation model is pre-trained on millions of synthetic tables. It has already learned the general shapes that turn up in tabular data, so it reads your labeled examples as context and predicts straight away.
In practice: seconds instead of weeks, and no penalty for changing your mind about the criteria.
Classification or regression
Text or few-valued targets become classes. Continuous numbers become a regression. Detected automatically, and you can override it.
Confidence on every row
Classification returns class probabilities, so you can sort by certainty and review only the rows that need a human.
Messy CSVs are fine
Missing values, mixed text and numbers, columns that appear in only one file. The app reports exactly what it used.
What you get back
One column appended
Nothing else changes. Row order, column order, and original values stay as they were.
| account | arr | tickets_90d | plan | prediction | prediction_confidence |
|---|---|---|---|---|---|
| Halcyon Group | 48,000 | 2 | enterprise | retain | 0.93 |
| Bellweather Co | 9,600 | 17 | growth | at risk | 0.88 |
| Orchard Labs | 22,400 | 6 | growth | retain | 0.61 |
| Vantage Print | 3,100 | 24 | starter | at risk | 0.95 |
Where it fits
Anything you can label
Lead and account scoring
Label the accounts that closed. Score the rest of the list against the same bar.
Churn and renewal risk
Label who left and who stayed. Flag the accounts heading the same way.
Review triage
Label the cases that needed a look. Rank the queue by how likely each one is to matter.
Data quality flags
Label the rows you know are wrong. Find the ones that look like them.
Category fill-in
Tag part of a catalogue by hand. Let the model finish the rest.
Estimates and grades
Continuous targets work too: price, duration, a numeric quality score.
Honest limits
What it will not do
The base model works best up to roughly 10,000 training rows, 500 columns, and 10 classes. Beyond that the app subsamples and tells you, rather than pretending.
It needs labels that mean something. Fifty carefully judged rows beat five hundred careless ones, and no model recovers a criterion you have not been consistent about.
It reads the columns you give it. If the signal is not in the table, it is not in the prediction.