GenAI for Ads Manager

My Role

Lead Product Designer

Strategic Integration

Directed the migration of experimental ML models into Meta’s primary revenue engine, Ads Manager. Navigated complex technical trade-offs to ensure GenAI features maintained the high performance and low latency required for a global advertiser base.

Human-in-the-Loop Governance

Developed the foundational frameworks for model output quality, including annotation rubrics and prompt engineering guidelines that ensured brand safety and aesthetic standards for global advertisers.

Governance & Trust

Established the "Advertiser Control" philosophy, ensuring AI-generated content adhered to strict brand-safety guidelines while maximizing creative flexibility.

Models released in Ads Manager

These models were tested on our experimental tool, AI Sandbox.

Text variations

Challenge

Advertisers struggle to create enough text variations, and they often use long text in ads. Today, more than 50% of the text copy has more than 20 words, creating digestive issues. Limited ad text liquidity and digestibility issues lead to suboptimal performance.

the solution

Improve ad performance by generating text variations that improve digestibility and diversity to reduce creative fatigue.

Image expansion

Challenge

Advertisers use a single image asset for multiple placements, which often results in ads that don't appear native to the platform - 60% of ads in IG Stories are not in the 9:16 format 

the solution

Eliminating Placement Friction. With 60% of assets lacking native 9:16 aspect ratios, I architected a GenAI solution that programmatically extends assets, ensuring a seamless, high-fidelity native experience on IG Stories and Reels.

Background generation

Challenge

Advertisers don't have the budget and expertise to create enough creative variations, so they limit their creative's reach to only audiences they think will resonate with their assets.

the solution

Use Generative AI to generate variations of the original images. This will enhance ad performance by increasing liquidity and avoiding creative fatigue.

Case study

Interactive segmentation

Advertiser facing GenAI features for background generation

Challenge

Images uploaded by advertisers into AdsManager are complex with cluttered backgrounds and often confuse AI models, leading to inaccurate automatic product extraction.

Solution

Integrated the Segment Anything Model (SAM) into a simplified user interface, allowing for manual precision where automation fell short.

Impact

Achieved a 56% effective segmentation rate and a 7% increase in overall feature adoption, proving that user control is the primary driver of AI trust.

Strategic Outcomes

Trust as a Metric

By balancing AI autonomy with granular user intent, we drove a 15% increase in adoption for background generation.

Global Scale

Successfully deployed these capabilities to 4 million+ advertisers, transforming GenAI from a "novelty" into a foundational creative workflow within Meta’s ecosystem.

How we tested these hypothesis in an experimental tool

The models seen in this presentation were part of the AI Sandbox.

View case study