Real engagements, anonymised. The problem, the approach, and what actually changed.
A churn model that catches three times as many real leavers as a random pick, tested only on months it never trained on.
On the reference week, the optimiser found €5,727 and 1,701 km of savings, with one truck fewer and no broken delivery appointments.
Gas churn concentrates in winter. A model that separates a price problem from a cold month, tested on a winter it never saw.
In telecom, timing is in the operator's own hands. A model that finds who will use the exit window — and the honesty that a proactive call can backfire.
Insurance churn happens at renewal. A score that may target attention but never price, tested on a renewal cycle it never saw.
In banking, leaving almost never means closing the account. A model built on the relationship, not the account, tested on months it never saw.
A two-sided health-creator marketplace built gate-first: approval before visibility, a badge only an administrator can set, and a brief that reaches only creators still approved the moment it is sent.
A clinical workspace where software reads a document once, a clinician confirms what it found, and the result is reused across timeline, registry, dataset and draft report.
A real-world-evidence platform where one definition sits below every dashboard, export and AI answer — so last quarter's figure still reproduces today, and no privileged path reaches a direct identifier.
A weekly regional-prioritisation platform that tells a marketing and sales team where to focus, why, and through which dealer — with the state of every data source shown next to the score.
