AI & data science

What we measure in AI work is not model accuracy but decision improvement. A highly accurate model that changes no decision produces no value.

Decision first, model second

The first question we ask on a project: which decision will this model’s output change? Does a prediction become an order, a campaign, or a stock decision?

Models built before answering that question usually work well and go unused. Once you know where the output connects, you also know how accurate the model needs to be — and it is often less than assumed.

The problem types we work on

Forecasting. Predicting product demand daily, weekly or monthly. In retail this is the input to stock level and shipment decisions.

Classification. Churn prediction, review sentiment analysis, problem categorization. In our Comentio product, marketplace reviews first pass through sentiment analysis, then the negative ones are classified by which problem they concern.

Segmentation and value. Customer segmentation and lifetime value prediction; the foundation of campaign targeting and recommendation systems.

Price behaviour. Price elasticity models: the effect of a given price change on sales and margin. This is the first phase of the DISCO product.

When it meets optimization

A forecast on its own produces a suggestion; combined with optimization it produces a decision.

A demand forecast is the input to the model that decides how many units go to which store. A price elasticity model is the input to the optimization that decides which product gets which discount in which week.

Building these separately is possible, but building them together produces fewer errors: the uncertainty in the forecast can feed directly into the constraints of the optimization.