Machine Learning Engineer, Churn Models

Churnora's predictor is a behavioral model running in our own code, and it has one job: rank a book of customers by how likely each is to leave, for reasons a person can check. Today it is a calibrated logistic regression over eight engineered features, chosen because it is inspectable. Your work is to make it better without making it a black box.

The bar is not cleverness, it is trust. A customer-success lead acts on our scores in front of their own customers. A factor that cannot be traced back to a number in that account's data is a bug, whatever the AUC says.

What you would own

• Own the scoring model end to end: feature engineering, fitting, calibration, and the evaluation set every change is measured against.
• Build per-tenant refitting, so a Scale customer's model learns from their own labeled history rather than our bundled cohort.
• Keep every factor honest: each one the product shows must trace to an observable number in that customer's row.
• Publish the method. The How the model works page is generated from the code and it stays that way.

What we are looking for

• You have shipped models that ranked or scored things in production, and lived with the consequences.
• You reach for the simplest model that answers the question, and you can say out loud what a coefficient means.
• You write TypeScript or are happy to; the scorer runs in the product, not in a notebook.
• You think a held-out metric is the start of the conversation, not the end.

How hiring runs

A thirty-minute call with a founder, then a paid work session on a real problem from our backlog, three to four hours, then conversations with two people you would work alongside. Four steps, usually inside two weeks, and we tell you where you stand at every one.