Accelerating Creative Testing with AI

The Challenge

The Marketing team relied on in-person testing to get feedback on creatives. That process often took weeks and could slow campaign launches significantly. We needed a faster, scalable solution to measure and predict creative performance.

The goal: Build a platform to predict creative success in hours instead of weeks, with measurable accuracy.

My Role: Building ASAP Auditor

As part of MediaFuturesGroup in media activation tech, I built the earlier version of ASAP Auditor, an AI-powered platform to measure and predict creative performance. The system was built on Vision AI, a Google Cloud technology that analyzes and understands visual data.

Technical Implementation

I developed the foundational infrastructure from the ground up:

Data Pipeline Architecture

Built the ingestion pipeline to collect and store creative performance data at scale that stores the data into BigQuery
Developed dbt (data build tool) models to transform raw creative metrics into structured datasets
Established data quality checks and validation workflows to ensure reliable model training

Vision AI Integration

Implemented text extraction from creative assets to analyze copy density and messaging
Built brand element positioning and prominence detection to measure brand visibility
Developed entity extraction and detection to identify significant objects and elements in images
Created logo detection and verification checks to ensure brand consistency
Built the API integration layer for processing image data and extracting these creative variables
Developed the Vision API call infrastructure to handle batch processing of creative assets

Predictive System Foundation

Created the architecture for correlating visual features with performance outcomes
Built data models to identify key creative variables (color schemes, layout composition, text density)
Established the framework for training predictive models on historical creative performance

How It Works

The platform I built analyzes creative assets using Vision AI to extract quantifiable features:

Text extraction to measure copy density and messaging effectiveness
Brand element positioning and prominence detection for brand visibility metrics
Entity detection to identify significant objects and visual elements in images
Logo detection and verification to ensure brand consistency across creatives
Text-to-image ratio analysis
Visual composition metrics

This foundation enabled the system to predict creative success, reducing testing time from weeks to hours.

Technical Stack

Data Infrastructure:

BigQuery for data warehousing
dbt for data transformation and modeling
Python for pipeline orchestration

AI/ML Components:

Google Cloud Vision AI for image analysis
Custom feature extraction models
Performance correlation frameworks

The Impact

The initial implementation I built demonstrated:

Automated creative analysis at scale across multiple campaigns
Rapid testing cycles without traditional focus groups
Data-driven creative optimization based on visual features

This version formed the basis for a system that could predict creative success with meaningful accuracy, fundamentally changing how Google’s marketing team validated creative performance.

Key Technical Learnings

Building this system from the ground up highlighted several critical considerations:

Data quality and consistency are foundational for reliable AI predictions
Vision AI feature extraction requires careful calibration to match business performance metrics
Pipeline architecture must be designed for iterative model improvements
Early prototypes need flexible design to accommodate evolving requirements
BigQuery and dbt provide a powerful combination for ML data preparation

The infrastructure matters. The data pipeline, API integrations, and processing architecture you build determine how effectively the AI system can scale and improve over time.