A model doesn't get to publish until it earns it.
Demiton fits models on your own history and writes their predictions into the same record as everything else. Before a model writes a single one, it is scored on history it never saw. If it doesn't clear the bar, it doesn't publish.
Models are on every plan, and run on the plan's agent and model hours: 10 a month on Insights, 80 on Connected.
Two models ship today.
Win probability
The chance of winning each open opportunity, learned from the ones you won and lost: the value, the customer, the type of work and how long the sale ran.
Margin drivers
How much each project's margin moved with cost growth, labour, delay and variations, so the next price starts from what moved the last one.
Or describe your own: what you want predicted, and what it should learn from. Demiton builds the dataset from your registers and scores it against the same bar.
The bar every model has to clear.
Scored on history it never saw
Several expanding folds: train on everything before a date, score what came after, step forward, repeat. Not one split that happened to land well.
The spread, not just the mean
If the folds disagree by more than 0.12, the verdict is unstable. An average of scores that disagree is not a measurement.
A floor it has to clear
Below an AUC of 0.60 (0.50 is a coin flip) a model does not write a single prediction. A model that predicts a number has to clearly beat predicting the median.
Too good reads as a leak
Above 0.95, Demiton assumes an outcome has slipped into the inputs. Operational data rarely allows near-perfect prediction.
No tuning to flatter it
No hyperparameter search, no threshold fiddling. Those make a weak model look strong without making it any better.
Reasons, not just a score
Every verdict says why in plain English. Usually it is a feature that reaches too little of the data, or too few outcomes to learn from.
Start with the history you already have.
The more jobs, opportunities and outcomes your registers hold, the more a model has to learn from. If there isn't enough yet, the verdict says so.