Project Overview

Company: Google
Role: Project vision / Lead designer
Team: 1 Designer, 1 ML researcher, 4 Eng, 1 project manager
Timeframe: 6 months

I initiated the project Milo to find a balance between design quality and performance optimization through machine learning in ad formats.

I worked with a small engineering team to create a new ad format from scratch that integrated machine learning tweaks into the template design in a controlled manner.

This approach aimed to maintain high visual quality while still leveraging the performance-enhancing capabilities of AI.

Challenges and Initial Context

Traditionally, ad formats are designed as responsive template with no moving parts other than the advertiser provided assets (text and images).

Initially, engineers were using machine learning to tweak existing ad formats with no designer involvement, leading to performance-driven but often visually poor outcomes.

This created tension between the engineering and design teams as designers were often left in a position of damage control.

My goal was to create a project where machine learning features would be baked in the project from the start. Designers could setup predefined boundaries to preserve design integrity while providing room for the model to explore and run experimentation without sacrifying user experience or product quality.

Design Strategy and Collaboration

My idea was to create a new type of ads format built from several UI elements that machine learning models could pick from and experiment with in A/B testing. This would allow some freedom for the model to fine tune its understanding of the ad format while giving the designer control over the outcome. For instance, the model will have a selection of button styles, fonts, color strategies, backgrounds...

The designer can take ownership of these permutation and be in the driver seat of the machine learning models rather than being subject to it. Designers will be able to preselect the UI elements and establish rules on how the model could use them (for instance by avoing certain combination, prioritizing button size over headline...).

To kick things off, I began by identifying common adjustments made by engineers and analyzing patterns in ad performance.

Next, I designed a new ad format system with flexible UI elements, accompanied by a set of rules and best practices. The system was built from the ground up to be easily scalable, allowing other designers and teams to seamlessly integrate it into their workflows.

To bring this vision to life, I collaborated closely with a small group of engineers to develop a working prototype and craft a clear, compelling vision for the project. Storytelling and internal marketing played a critical role in securing buy-in from both the engineering and UX teams to ensure the project's success.

Navigating Constraints and Experimentation

One of the main challenges was aligning the internal machine learning infrastructure with the project's requirements.

The existing models were not equipped to handle complex rules and constraints, leading to delays as we worked to prioritize changes.

To address this, I consolidated all the necessary rules and permutations into a single comprehensive request, allowing the back-end team to implement them efficiently.
This approach minimized disruption and helped the project progress despite limitations.

Impact and Outcomes

The first set of experiments revealed that fewer permutations were needed for effective optimization and that high visual quality did not always correlate with increased performance.

This led to a reframing of the project as a premium ad format suited for brand-conscious advertisers who prioritize quality over raw performance.

The success of this approach demonstrated that a balanced, controlled use of machine learning could achieve both high-quality design and strong performance metrics.

Key Takeaways

The Milo project highlighted the importance of collaboration and strategic compromise between design and engineering.

By creating a structured framework for machine learning, I was able to establish a middle ground where both quality and performance could thrive.

The project also underscored the power of strong visual storytelling and realistic mock-ups in gaining buy-in and shaping the direction of product development.