How Cross Screen Media Trains Models for 210 Markets in Hours, Not Weeks, with ZenML

The data science team at Cross Screen Media manages ML workflows across 210 local markets on trillions-of-rows-scale tabular data. Through two key projects, they demonstrated ZenML's impact: first, a critical pipeline that previously took ~1 week now completes all 210 markets in ~2 hours after code optimization and ZenML implementation. Building on this success, the team applied ZenML to a second project involving model development, achieving a ~17% average AUC improvement across all markets. These wins freed the 3-person team to tackle more ambitious methodologies.

The Challenge: A Week-Long Training Bottleneck

At the core of Cross Screen Media's business is its ability to plan and measure hyper-local advertising campaigns across 210 markets. Working with massive structured datasets scaling to trillions of rows, the data science team of three found themselves spending more time managing infrastructure than building models.

This manual process created several critical issues:

• A Two-Codebase Problem: The team was forced to maintain two separate codebases—Jupyter notebooks for experimentation and different Python scripts for production. This required constant, error-prone translation between environments.
• Operational Risk: The workflow created single points of failure. As the team noted, if the person responsible for a model run was sick, the entire process could stall, jeopardizing delivery.
• A Week-Long Bottleneck: Running models for all 210 markets was a slow operation that stifled iteration and took over a week to complete. The hardware and process limitations also forced them to simplify their models, sacrificing predictive accuracy for the sake of finishing the job. The team knew they could build better models but lacked the platform to do it efficiently.

The Solution: A Standardized Platform for Automation

The team chose ZenML to build a new MLOps strategy. With ZenML, they engineered a standardized workflow, building a single codebase with pipeline templates that could be parameterized for each of the 210 markets.

This change meant a model could be tested locally on a small data sample with confidence that it would scale in production. ZenML acted as the orchestration layer, empowering the data science team to operate as a self-sufficient unit

This investment in standardization pays dividends for future work. "The incremental cost of new projects is much less now," Shey and Junyi Lin add. "Once you have the patterns and templates set up, you can reuse most of it for subsequent projects."

A Partnership That Accelerates

Beyond the technology itself, the CSM team highlighted the value of ZenML's responsive support model. In contrast to experiences with larger providers, they noted the benefit of having direct access to the technical team. "The team has been responsive and quick to work on the unique challenges we faced," Shey and Junyi Lin noted, ensuring that implementation hurdles were overcome without losing momentum.

The Results: Two Projects, Two Wins

Cross Screen Media's ZenML implementation delivered impact across two sequential projects, each demonstrating different dimensions of value.

Project 1: Pipeline Acceleration

The first project focused on optimizing a critical data pipeline that processed all 210 markets. The team refactored their code and implemented ZenML orchestration, collapsing the runtime dramatically.

Runtime collapse: A pipeline that previously required ~1 week to complete across all 210 markets now finishes in ~2 hours. (This reflects both ZenML orchestration and code optimizations.)

This acceleration was transformative. The team no longer spent days waiting for pipeline runs to complete, freeing them to iterate and explore more sophisticated approaches.

Project 2: Model Quality Improvements

Building on the infrastructure and patterns established in Project 1, the team applied ZenML to a second initiative focused on improving their modeling methodology. With the platform handling orchestration and parallelization across 210 markets, they could invest in more advanced techniques.

"ZenML gave us the opportunity to make our models more complex and add more attributes," said Shey and Junyi Lin. "This was impossible before because we lacked the right platform."

The additional model complexity paid off. By implementing more sophisticated approaches, the team achieved a ~17% average improvement in AUC scoreacross all 210 markets—a direct enhancement of Cross Screen Media's core product quality.

Operational Excellence

Beyond the project-specific wins, the ZenML dashboard became the team's command center for MLOps, providing observability across all pipeline runs.

• Fewer kubectl dives: The team now debugs primarily in the ZenML UI; shelling into Kubernetes is rarely necessary.

This centralized observability provides long-term business value. The ZenML dashboard acts as a historical ledger, allowing the team to look back at year-old model runs, see which markets failed, and track performance metrics over time—a capability that was impossible with their previous, fragmented approach.

Looking forward

The team is eager for fully dynamic templating of market subsets to reduce manual grouping. This aligns with ZenML’s ongoing templating work and will further simplify large-fanout runs.

By solving their primary bottleneck of speed, the team at Cross Screen Media unlocked a new level of model performance and business value.