March 7, 2024 
Picture a Monthly Review Meeting in a Fintech.
Can be stressful!
Has there been growth in the Monthly Active Users (MAUs)? Is Customer Acquisition Cost (CAC) trending downwards? How is the Retention Rate (RR)? And so on and on.
The Product Managers and the CXOs are in a state of stress if these numbers are not exactly in rude health.
They go,
“Our MAUs have plateaued for some quarters, and other metrics are also under stress. We are missing that personal touch that resonates with our users. Currently, we’re less able to offer financial products that truly adapt to our users’ unique financial situations.”
“A tailored savings account where the interest rates and saving recommendations are not one-size-fits-all but instead dynamically adjusted based on the user’s spending habits, financial goals, and even lifestyle choices. It could encourage more savings by showing users exactly how altering their spending habits could directly impact their financial future.”
“For example, imagine if we could offer not just generic financial advice but personalized investment strategies that adapt to each user’s financial goals, risk tolerance, and historical investment patterns.”
“Each user gets a unique experience, feeling like our platform truly understands and caters to their individual needs. And I think that would help push our metrics up!”
“A customizable insurance product that dynamically adjusts coverage and premiums based on real-time data from the user’s life events, health metrics, and financial changes. These aren’t just improvements; they’re game-changers.”
Then it happens! They turn towards the CTO and ask “that” question:
“All this talk of GenAI – what could you do to help us?”
The CTOs are, of course, absolutely expecting these questions these days!
“Guys, your vision for hyper-personalization aligns with the capabilities of Generative AI (GenAI). GenAI can analyze vast amounts of data to provide customized financial solutions, enhancing user engagement, and potentially boosting our MAUs and other metrics.”, says our CTO.
She continues, “Here is what we can do – nothing big and shiny – let me just give you a sense of what is possible over the next 4 weeks. And, yes, before you ask, preliminary results should be with us within 6-8 weeks post-implementation.”
Goal: Leverage Generative AI (GenAI) to enhance Monthly Active Users (MAUs) and Average Revenue Per User (ARPU). What could we expect?
- a personalized model could increase customer conversion rates by 8-15%, reduce churn by 20-30%, and grow ARPU by 12-18%
- Set an internal goal, based on current baselines, to increase MAUs by 10% and retention rate by 25% over 6 months post-launch
Investment choices: We will clearly focus on balancing performance with computational costs to ensure that the project remains financially viable, with ROI anticipated through improving business metrics over time. Need £40K to £100k – a breakup follows in Annex 1 with rationale.
4-weeks Implementation Plan
Foundational
GenAI Technologies
- We will focus on Retrieval Augmented Generation (RAG) for dynamic, data-driven personalization.
- We will use an auto-regressive transformer architecture like GPT-4 for next-token prediction and language generation. This would form the core personalization engine based off a causal understanding of user preferences.
- Additionally, we will incorporate non-causal bidirectional transformer layers akin to BERT to absorb larger context and high-dimensional data. This allows accurate situational predictions informed by a rich set of historical data points.
- Team to further optimize using Proximal Policy Optimization (PPO), a model-free reinforcement learning algorithm suitable for financial recommendation tasks with discrete actions and stochastic transitions.
- The key benefit from this hybrid architecture? Combining the adaptability of causal models like GPT that can dynamically generate personalized responses for each user, with the enhanced contextual representation learned via techniques like BERT.
- The RL training also directly optimizes sequences of financial recommendations for long-term returns.
- The resultant model would surpass relying solely on pre-trained models like GPT-4 by specializing to the fintech domain with custom techniques for personalization and optimize risk-adjusted performance.
- Technical KPIs Model inference time, accuracy, and computational costs
Data Strategy and Model Selection
Data Engineering:
- We will implement advanced preprocessing, utilizing differential privacy and federated learning for data anonymization without compromising utility.
- Team to source transaction graphs to model interconnected financial behaviours vs simplistic attribute/event data
- Next to think about generating privacy-preserved synthetic data mirroring distributions in actual user data, curating niche domain-specific datasets like location-based spending habits to enhance personalization, etc.
- We would then implement active learning for users to directly improve their models through feedback loops, and ensure secure handling to meet GDPR and CCPA standards.
Model Selection: team actions would be
- Tailor RAG models with financial domain data for insightful advice.
- Use TensorFlow Decision Forests for adaptive predictive analytics, ensuring cost-efficiency in training and runtime.
- Establish concept drift triggers based on deteriorating model inference-time performance
- Maintain human-in-the-loop oversight through mechanisms like LIME model explanations
- Containerize models for A/B testing experimental models on subgroups
- Establish automated CI/CD pipelines with model regression testing
- Infrastructure as code for cloud provisioning and scaling
- Implement version controlled data and model registries
- Use online learning and shadow modes to test models pre-deployment
Integration and Prototype Testing
Integration:
- Ensure API adaptability with existing systems.
- Utilize AWS Lambda for scalable deployments, focusing on secure, compliant data pipelines.
Testing:
- Develop a prototype; conduct A/B testing.
- Use TensorFlow Model Analysis for in-depth performance insights across user segments.
- Expose model to 5% of users initially for quantified analysis
- Study changes to platform KPIs and individualized business metrics versus holdout group
- Only proceed to further 20% rollout after predetermined targets are hit
Iteration, Deployment, and Scaling
Iterative Improvement:
- Refine models based on prototype feedback, optimizing for real-time performance and cost-efficiency.
Deployment Strategy:
- Implement blue-green deployment to minimize user disruption.
- Apply automated scaling and continuous monitoring to support growth.
- Allocate 2-3 weeks for intensive small-scale testing versus 1 week in MVP
- Carefully segment users into homogeneous batches based on key attributes to isolate issues
- Enforce caps per segment for maximum allowed errors before rollback
Cost Management and Monitoring:
- Leverage cloud cost optimization tools to maintain budget control.
- Establish metrics for ongoing cost-performance evaluation.
Annexure 1
Here is the broad break-up:
| Cost Category | Estimated Range (£) |
| Cloud Computing Costs | Utilizing services like AWS Lambda for model deployment and inference can lead to variable costs based on usage. Assuming a moderate level of API calls and computational requirements, costs could range from £500 to £3,000 monthly. |
| Development Team* | Assuming a 4-week sprint with a team of 4-5 members, costs could range from £30,000 to £50,000, considering the UK’s average salary/contract rates for such roles, also considering the high demand for these roles in the current GenAI skills’ demand context. |
| Software and Tools Licensing | Depending on the choice of tools and platforms (outside of open-source options), licensing for specialized software for data processing, development, and deployment could add an additional £1,000 to £5,000. |
| Data Acquisition and Processing | If external data sources are required or if significant costs are incurred in preprocessing large datasets, this could add £1,000 to £10,000, depending on the data’s complexity and volume. |
| Security and Compliance | Ensuring GDPR and CCPA compliance through security measures such as data encryption and access control could add £5,000 to £15,000 to the initial setup costs, considering consulting fees and implementation of security protocols. |
| Testing and Deployment | Costs for A/B testing tools, prototype development, and deployment strategies such as blue-green deployment could range from £2,000 to £10,000. |
Development Team
| Role | Number of Individuals | Skills Required | Responsibilities |
|---|---|---|---|
| Data Scientists | 2-3 | Expertise in GenAI technologies (e.g., RAG, GPT-4), machine learning, deep learning, and data analytics. Familiarity with TensorFlow, PyTorch, and other ML frameworks. | Model development, training, and fine-tuning. Data analysis and insights generation. |
| Machine Learning Engineers | 2 | Proficiency in implementing, scaling, and deploying AI models in production environments. Experience with AWS Lambda, Docker, and Kubernetes for deployment. | Operationalizing models, ensuring scalability and efficiency. Integration with existing systems. |
| Data Engineers | 1-2 | Strong background in data architecture, ETL processes, and handling big data technologies (e.g., Hadoop, Spark). Knowledge of data privacy practices. | Data preprocessing, cleansing, and ensuring data quality. Implementing data privacy and security measures. |
| Software Developers | 2 | Experience in backend development, API integrations, and frontend development for prototype interfaces. Familiar with agile development practices. | Building the user interface for the prototype, integrating AI outputs into user-facing products. |
| Project Manager | 1 | Strong leadership, communication, and project management skills. Experience in tech projects, particularly AI or fintech. | Overseeing project timelines, budget management, coordinating between teams, and stakeholder communication. |
| Security and Compliance Officer | 1 | Knowledge of data protection laws (GDPR, CCPA), cybersecurity practices, and fintech regulations. | Ensuring the project meets all legal and compliance requirements. |
Caution:
While the regulatory environment in the UK does support innovation and the personalization of financial products, such as those mentioned above, fintechs must navigate these regulations carefully. They must ensure complying with the FCA’s rules and principles, particularly the new Consumer Duty, to provide products and services that are not only innovative but also fair, transparent, and in the best interests of their customers.
Read More