Boris Klaydman
WorkCloudKitchens

An in-house analytics studio that cut churn 8%

A self-serve analytics platform for a company that runs on data. Datasets with live lineage, dashboards anyone can assemble, and prediction features that told operators what was about to happen instead of what already had.

Client
CloudKitchens
Role
Senior Manager, Design and Product
Year
2022
8%reduction in customer churn
17%fewer support calls after the generative AI support flow
23%more orders from streamlined onboarding

The problem

CloudKitchens runs thousands of delivery-only kitchens on operational data. The data existed. The ability to see it did not. Teams queued behind analysts for every question, and operators found out about problems after they had already cost money.

What we built

Analytics Studio: datasets, metrics, dimensions and filters as first-class objects, with live lineage that shows where every number comes from and what depends on it downstream. Operators assemble their own boards. Analysts govern definitions once instead of answering the same question forever.

The metrics index: every metric with a health score, tier, certification and an owner. Governance as a visible property, not a wiki page.

A dataset as a first-class object: Customer Orders with its size, run cadence and cost, and everything that depends on it fanned out to the right

The same dataset from the lineage side: one source feeding 172 metrics, 19 boards and their consumers

The lineage view

The trust problem in analytics is always the same question: where did this number come from? Lineage answers it in place. From any metric, you can see the datasets it draws from and every board it feeds, and click through the chain in either direction.

While on a dataset page, the user clicks through dependent entities to trace the lineage relationships

A metric’s full fan-out: one definition of Paying Occupancy Rate feeding 172 metrics and 19 boards, with the Otter Orders board opened in place

The lineage graph stays live while editing: config, squad and activity are one tab away from the graph itself

Downstream views compressed: how the lineage model was storyboarded, from initial state through metrics trees to board and metric graphs

On top of it we built predictive consumer traffic analytics and recommendation features that warned operators before demand shifted, and a generative AI customer support flow that resolved common questions without a call.

Impact

Churn fell 8 percent. Support call volume fell 17 percent. The streamlined multi-service menu onboarding that fed the same data raised orders 23 percent.