How a six-person data team serves 1,000 people at Atolls

Atolls grew largely by acquisition, and every acquisition brought its own data stack. The data team inherited data spread across Redshift, two separate Athena instances, and ClickHouse, and had to answer questions that cut across all of them. Today, the team of six works in Deepnote: a single workspace where they query every data lake side by side, publish interactive analyses and self-serve data apps for stakeholders, and run their first machine learning models in production.

Atolls' business model means regularly buying other companies, and with each one comes another data lake. Before those sources are migrated, the data team still has to compare user bases and understand how each business is doing across systems that were never designed to talk to each other.

That team sits at the center of the company. Its stakeholders are marketing, product, and the C-level, and its work runs from campaign analysis and audience extracts to segmentation and predictive modeling. A BI tool covers standard dashboards; everything deeper happens in notebooks.

Every acquisition brought another data lake

Answering a single question could mean pulling from Redshift and Athena separately: write a query in one IDE, download a CSV, query the next source, download another CSV, then stitch the files together in a Jupyter notebook and hope the extracts were right. The friction was bad enough that some analysts ended up specializing in just one data lake.

Analyses lived in Confluence and aged badly

Once the analysis was done, the team exported plots as images and wrote everything up in Confluence. Any stakeholder feedback meant going back to the notebook, rerunning it, often discovering the environment had changed in the meantime, recreating the plots, and overwriting them in Confluence. Getting a finished analysis in front of stakeholders took one to three days, and each feedback round could add two or three more days of work.

1 day a week went to data extracts

Marketing regularly needed audience extracts: users matching certain criteria, pulled to a CSV. Each request meant an analyst writing SQL by hand. In high season, with lots of campaigns running, these simple requests were eating one to two days of the team's week.

Every data source in one notebook

In Deepnote, the team has connected Redshift, Athena, ClickHouse, and S3 to the same workspace and queries them in the same notebook: two SQL cells, one workflow, SQL and Python mixed as needed. Comparing user bases across acquired businesses no longer requires CSV archaeology, and analysts are no longer boxed into a single data lake.

Shareable links instead of Confluence pages

Analyses now live in Deepnote and are shared with stakeholders via a link, with narrative in Markdown and interactive Plotly charts instead of static screenshots. Everything stays stored in one place, so stakeholders can always come back to past work. When feedback comes in, the team updates the notebook and the analysis is current again, no re-exporting, nor overwriting pages.

Self-serve data apps for the entire team

The team runs about five data apps in regular use. Orgs like marketing use them as an interface to the company's data: pick the criteria, generate the audience, download the CSV - no ticket to the data team required. The whole marketing department uses the apps regularly, and the recurring extract requests have essentially disappeared.

AI that knows the context

The team has used Deepnote agent from the start, and they've become part of daily work: fixing code in place, autocomplete, and cell-level AI assistance, with a choice of models. Because the AI has the context of the whole analysis, nobody copy-pastes code into an external chat anymore.

From three days to same day

An analysis that used to take one to three days to reach stakeholders now goes out the day it's finished. Feedback that once triggered days of rework now takes at most a day, often hours, sometimes minutes.

8–10 hours a week back

The data apps automated the recurring extract requests that used to consume up to two days a week. That time went back into analysis and new data products.

The team's first ML models in production

The team had wanted to ship machine learning models for a while, but the deployment options they evaluated all demanded too much setup. Deepnote gave them a way to start: a notebook and lightweight ETL in one, running on schedules. Today they operate a customer segmentation model, a customer lifetime value model, and a churn prediction model in development. They even built an in-house review agent: scheduled notebooks gather product reviews from Trustpilot and the app stores, an internal LLM categorizes the issues, and the results feed a dashboard and suggested responses.

Deepnote as the agent workspace for the team

Adoption took some iteration. The team learned to design apps for non-technical users, and users occasionally surprised them, like the stakeholder who discovered he could compare 20 campaigns at once in an app built to benchmark one. But the shift happened.

With their new enterprise plan, the team is rolling out Deepnote beyond the data team, giving colleagues across a company of about 1,000 people secure access to analyses and apps via single sign-on.

The bigger ambition is rethinking what a data team looks like in the AI era. Stakeholders increasingly explore data with LLMs on their own, and the Atolls team sees its role evolving: building tailored data apps, shipping ML models quickly, running automated data-quality checks, and verifying that the insights stakeholders get from AI are actually legitimate. The team is also exploring Deepnote's MCP support to connect their notebooks with Claude.