How Comcast Advertising Accelerates Campaign Forecasting with Databricks Apps
Comcast Advertising connects brands to nearly 125 million U.S. households through multiscreen TV campaigns spanning traditional and streaming platforms. The company’s data science team used Databricks Apps to transform complex predictive models into interactive forecasting dashboards accessible directly by sales and marketing teams. Development cycles shortened by 10–30%, with campaign managers gaining the ability to run what-if scenarios in real time without data science support.
Impact
10–30%
Reduction in data product development time
125 million
U.S. households reachable through Comcast Advertising campaigns
Challenge
Data scientists built accurate predictive models for campaign optimization but had no practical way to deliver interactive model outputs to business users — custom application development required separate hosting and unfamiliar front-end tooling, slowing the path from insight to decision.
Solution
Comcast Advertising deployed Databricks Apps to build Python-based interactive forecasting dashboards directly within their existing Databricks platform, using Unity Catalog for access control, MLflow for model serving, and SQL Serverless for governed queries.
Tools & Technologies
What Leaders Say
“We wanted a way to let business users interact with model outputs directly, without requiring them to understand the underlying data science. That meant we needed a platform that let us quickly build applications, customize them and integrate with our data pipelines.”
“We’ve saved time, but more importantly, we’ve improved the quality of what we’re building by keeping everything within a unified platform.”
Sign up to read complete case studies, access detailed metrics, and unlock all use cases.
Full Story
Comcast Advertising operates at a scale that makes data science both indispensable and a bottleneck. Reaching 125 million U.S. households across all 210 designated market areas generates massive volumes of viewership, engagement, and campaign performance data. The organization’s data scientists built sophisticated predictive models to optimize ad strategies — but those models lived in technical environments that business users could not access or interpret.
The gap between model output and business decision was a costly one. Traditional BI tools lacked the flexibility to let end users interact with model outputs dynamically. Building custom applications for business teams required separate hosting infrastructure, unfamiliar front-end frameworks, and significant engineering overhead that had nothing to do with improving campaign performance. Data scientists spent time on application plumbing rather than model improvement.
When Databricks Apps became available, the Comcast Advertising data science team recognized an opportunity to close the loop. Already using Databricks for Lakeflow Spark Declarative Pipelines, ML experimentation, Unity Catalog governance, and MLflow model tracking, the team could now build Python-based interactive applications directly within their existing data platform. No separate hosting, no context switching, no new tech stacks to maintain.
The first application was a forecasting dashboard that lets business users adjust input levers — budget allocations, audience targets, campaign parameters — and see immediate revenue predictions. Sales, strategy, and customer experience teams can run unlimited what-if scenarios without queuing requests to data scientists. The feedback loop tightened dramatically, with model outputs reaching decision-makers faster and with more contextual interaction than any static report could provide.
Development time for new data products dropped 10–30%, a consequence of keeping everything inside a unified platform. The forecasting dashboard is currently in user acceptance testing, with plans to expand to hundreds of users. The team is also exploring generative AI capabilities within Databricks, including AI assistant chatbots and the Lakebase architecture to further unify operational and analytical data.