EASA
Ad Libra: AI-driven ad compliance monitoring at scale
Summary
An AI-powered advertising analysis platform designed to automatically screen thousands of online ads and assess their compliance with local advertising rules using GenAI and Computer Vision, achieving 94% precision.
With the support of Google and Meta, we built an end-to-end AI solution for the European Advertising Standards Alliance (EASA) to automatically monitor social media advertising for compliance with local and European advertising regulations. The platform went from concept to production in 3 months, exceeding all performance targets.
Services
Industry
The Challenge
Ad volumes keep growing, but monitoring tools haven't kept up.
Ad volumes continue to grow, making large-scale monitoring of online advertising increasingly difficult for self-regulatory organizations tasked with ensuring companies and brands comply with advertising rules and regulations. Staff were overwhelmed by fast-growing online ad volumes and proliferation of formats. The core technical challenge was detecting specific violations within a dataset characterized by extreme class imbalance with over 90% organic content and limited labeled data.
Problem 01
Growing ad volumes outpacing monitoring
- ▪Staff overwhelmed by fast-growing online ad volumes
- ▪Proliferation of online advertising formats
- ▪Existing commercial monitoring tools not tailored to regulatory needs
- ▪Safety, security, and IP concerns with third-party platforms
Problem 02
Extreme class imbalance in detection
- ▪Over 90% organic content in the dataset
- ▪Only 6,000 labeled examples, far too few for robust classification
- ▪Required precision above 90% to avoid overwhelming reviewers
- ▪Limited labeled data making minority class detection extremely difficult
The Solution
We designed an operational AI architecture built for speed, scale, and regulatory accountability, combining GenAI-driven data augmentation with precision-focused model training and a complete audit trail for every decision.
01
Risk-driven AI architecture
Partnered with Google and Meta to implement effective data ingestion via APIs. Implemented a comprehensive, risk-driven AI architecture for scalable compliance monitoring where every AI decision is logged with a full audit trail.
02
LLM-powered data augmentation
Used an LLM to generate confidence scores for over 14,600 unlabeled posts. By filtering on high-confidence predictions, expanded the robust dataset from 6,000 to 20,600 examples, turning data scarcity into abundance.
03
Production deployment and integration
Integrated production standards with thorough user research, going from concept to production in just four weeks. Containerized deployment delivers predictions in under 500 milliseconds with full experiment tracking.
Impact
From concept to production in 3 months, exceeding all performance targets with 94% precision and transforming how regulatory resources are allocated.
Precision
94% precision exceeded the 90% threshold, minimizing false positives so valuable human reviewer time is spent on genuine compliance issues rather than noise.
Model Robustness
0.83 F1 score demonstrating robustness even for rare edge cases in a heavily imbalanced dataset, with a targeted 30/70 ratio preserving rare violations.
Operational Transformation
The system filters 90% of organic content with surgical precision and presents only high-confidence violations for review, transforming how regulatory resources are allocated.
“The system has transformed how EASA and their members monitor ad compliance across European markets, letting legal reviewers focus on genuine violations.”
Director of Digital Compliance, EASA
Methods
An end-to-end AI approach combining data augmentation, precision-focused training, and containerized production deployment.
GenAI
Computer vision
LLM confidence scoring
Data augmentation
Clean architecture
Systems thinking
User story mapping
Agile
Containerized deployment
Experiment tracking
API integration
Metadata catalog
Solution architecture
Engineering
Project management