We build predictive models for companies that do not have a data science team. We also talk a fair number of people out of one, which is a strange way to sell a service and the reason clients believe the rest of what we say.

The question we ask first

What decision changes if the prediction is right? If nobody would act differently, a model is an expensive dashboard. If someone would — reorder sooner, call that customer, hold that shipment — then it is worth costing.

When a rule beats a model

More often than the industry admits. If your rule is "flag any order over $5,000 from a customer who has never bought before," that catches most of what a fraud model would catch, runs instantly, and never needs retraining or explaining. We will build the rule, measure how much it misses, and only then tell you whether a model is worth the difference.

Rules also have a property models do not: when one fires wrongly, anybody can read it and see why.

When it is genuinely worth it

When the pattern is real but too tangled to write down, and you have enough history for it to show up. Demand forecasting with two or three years of sales. Churn prediction with a few thousand customers who have already left. Lead scoring where you know which deals closed.

How we build them

Simplest model that works, measured against the rule it replaces rather than against nothing. A number you can watch each month so you know when it starts drifting, and a plan for what happens when it does.

Bring us the decision, not the algorithm

Tell us what you would do differently if you knew. We will tell you whether the data supports knowing.