The Bank’s digital marketing was suffering from ‘Content Fatigue’. Generic assets led to declining CTR, while rigorous compliance reviews created a 4-week latency that made agile A/B testing impossible.
The Challenge
1. Executive Summary
Client: A Global Systemically Important Bank (G-SIB).
The Strategic Problem: The Bank’s digital marketing was suffering from “Content Fatigue.” In a saturated digital landscape, generic, uninspiring creative assets were leading to declining Click-Through Rates (CTR) and rising Customer Acquisition Costs (CAC).
However, the primary bottleneck was not creativity, but Compliance. The Bank’s rigorous review process for financial promotions (governed by FCA/SEC rules) created a 4-week latency for every new campaign.
This lag made agile A/B testing impossible; by the time a variation was approved, the market moment had passed.
The Solution: Project Nexus is a bespoke “Content Intelligence” platform built on the AWS and Databricks ecosystem. Nexus uses Generative AI to create content and Computer Vision to “score” creative assets for compliance and performance before they launch.
Technical Architecture
2. Technical Architecture: A Lakehouse for Creativity
The architecture follows a Data Lakehouse pattern, unifying data engineering, ML training, and MLOps into a single governed substrate.
2.1 The Creative Feature Store
Traditionally, banks track who they targeted (demographics). Nexus tracks what was shown (the creative asset itself). We built a Creative Feature Store using the Databricks Feature Store.
- Visual Features: We used a pre-trained CLIP (Contrastive Language-Image Pre-training) model to extract high-dimensional embeddings from thousands of historical ad images.
- Semantic Features: We used NLP to extract sentiment scores, reading ease (Flesch-Kincaid), and “Urgency Keywords” from historical ad copy.
- Performance Features: These creative features were joined with historical performance data (CTR, Conversion Rate, Bounce Rate) in the Delta Lake tables to train a predictive XGBoost regressor.
2.2 Generative AI & The “Compliance GAN”
For assets that needed creation, we deployed a Generative AI workflow.
- Generation: Llama 3 (accessed via Amazon Bedrock for data privacy) generates 50 variations of ad copy for a specific product, fine-tuned on the Bank’s “Tone of Voice” guidelines.
- The “AI Auditor” (Compliance Filter): This is the most critical component for financial services. A separate, smaller BERT-based classifier acts as a binary “Judge.” It flags any copy that violates regulatory guardrails (e.g., missing disclaimers) before it ever reaches a human reviewer.
2.3 The Closed-Loop Feedback
The system utilises Thompson Sampling (a Bayesian Multi-Armed Bandit algorithm) for dynamic A/B testing. Instead of waiting 2 weeks for statistical significance, the Bandit algorithm dynamically routes more traffic to the better-performing creative variants in real-time.
Operational Impact & ROI
Based on the deployment of similar architectures in high-velocity financial environments, Project Nexus delivered transformative results:
- Engagement Uplift (CTR): +35% Increase in Click-Through Rate. By moving from “gut-feel” creative to “data-scored” creative, the Bank eliminated low-performing assets before launch.
- Customer Retention: +18% Uplift in Retention for At-Risk Segments. The personalisation engine identified “churn-risk” signals and automatically deployed empathetic “Appreciation” content rather than sales pitches.
- Operational Efficiency: Reduced Compliance Review Time from 4 Weeks to 3 Days. The “AI Auditor” pre-screened content, eliminating the backlog of “obviously bad” content.
This case study describes work undertaken by the founder of Altablack prior to the firm’s creation, presented here to illustrate the technical and strategic foundations of the practice.
