About the Customer
The client is a renowned private, employee-owned investment management firm with its headquarters nestled in the bustling financial district of New York. It manages equities, fixed income, private equity and hedge fund portfolios for institutional, retail and high-net-worth investors.
Key Challenge/Problem Statement
Before engaging with Vertical Relevance, the client faced unexpected losses and missed revenue opportunities in their Institutional business, primarily due to the difficulty in keeping track of relevant changes at their clients. The client’s business analysts had to sift through targeted news sources and conduct public internet searches to compile weekly reports highlighting critical changes in their clients’ profiles that might indicate either new business opportunities or the risk of attrition. This process was time-consuming, prone to human error, and not scalable.
Moreover, the lack of automation and technical solutions to streamline these efforts meant that the client was unable to capitalize on more advanced technological capabilities, such as generative AI, to optimize their operations. This deficiency in automation and AI utilization posed a barrier to business growth and sustainability, as the manual processes hindered the client’s ability to disseminate crucial information quickly and accurately to account teams and ultimately to their clients.
State of Customer’s Business Prior to Engagement
The client has over $450B in Assets Under Management, spanning Public and Private Markets. Its strategies include equities, fixed income, hedge funds/alternatives, real assets, private equity and private credit. Its Institutional business serves corporate and public sector pension funds, endowments and foundations, sovereign wealth funds and other institutions.
The client had dealt with missed revenue opportunities and unexpected attrition of Institutional clients in recent months. Most of these events were retroactively traced back to key changes that had occurred at the clients and reported on the internet or industry news sources. However, the difficulty in accessing, filtering and assessing relevant news items manually and generating actionable insights had prevented the client from taking timely action to pursue new business opportunities or prevent losses.
Technically, the client had no designs or solutions in place to alleviate this burden through automation. Their infrastructure was predominantly on-premises, which, while stable, lacked the flexibility and scalability afforded by cloud solutions. The on-premises implementation further compounded their challenges, as it did not leverage the benefits of distributed computing, such as improved data processing speeds and the ease of integrating advanced technologies like generative AI.
Moreover, the client had not yet ventured into the use of cloud services or generative AI within these processes, despite recognizing the potential benefits. They were aware of the transformative power of cloud computing and AI but had not taken steps to integrate these technologies into their workflow. This gap in their technical capabilities was a clear shortcoming, as it precluded them from harnessing efficiencies and insights that could be gained from modern, cloud-based AI solutions.
In summary, the client’s existing approach was marked by manual, time-consuming processes, a lack of automation, and an on-premises infrastructure that was not conducive to scaling or integrating with advanced AI technologies. These shortcomings highlighted a critical need for a transformative solution that could automate their newsletter generation process, reduce the manual labor required, and drive greater business value.
Proposed Solution & Architecture
Vertical Relevance was tasked with transforming the labor-intensive process of generating client-specific newsletters for a client owner into an automated, efficient, and scalable solution using Amazon Web Services (AWS) and generative AI. The client required an enhancement of their current manual system, which demanded considerable hours from business analysts to compile valuable client information from numerous data sources into weekly newsletters.
The proposed solution involved a multi-component AWS architecture designed to automate the ingestion of relevant data, processing the data, retrieval of crucial information, and distribution of targeted newsletters. The solution employed AWS services and generative AI to streamline the collection of data from subscribed newsletters, Salesforce client information, and internet searches, thereby significantly reducing the manual effort required by the business analysts.
At a high level, the architecture consisted of data sources such as S3 buckets for storing newsletter data, Salesforce for client details, and internet searches for additional news. These data sources were connected to AWS Lambda functions, which triggered processing and metadata extraction. The information was then ingested into AWS Aurora RDS, a cost-effective vector database chosen for this proof of concept, leveraging Amazon Bedrock’s knowledge base to simplify data ingestion. Furthermore, AWS Fargate hosted a Python application that orchestrated the weekly processes and generation of the newsletters.
Automation was set up to pull data from email newsletters and Salesforce, storing this information in separate S3 buckets. A weekly cron job was established to extract client data from Salesforce into a CSV format and place it into an S3 bucket. Upon data arrival in the S3 buckets, Lambda functions were triggered to extract metadata and store both the metadata and source files in an S3 bucket configure as Bedrock knowledge base data source.
A one-time setup Lambda function was used to create the necessary schema in the RDS database, allowing the knowledge base to ingest the data correctly. A Python application, containerized and running on AWS Fargate, was triggered by CloudWatch rules every Monday. This application retrieved client information, performed internet searches, and processed the results, invoking generative AI models to determine the relevance of content and generate targeted newsletters. The generated newsletters were formatted and stored in an S3 bucket, with notifications sent to inform stakeholders of their availability. The S3 bucket was then synced to the client’s on-premises shared folder for easy access by the sales team and business analysts.
Terraform Cloud was used to manage infrastructure deployment securely, while Azure DevOps handled the Git repository and CICD pipeline, ensuring that the latest application versions were deployed seamlessly using updated Docker images in ECR and ECS task definitions.
Key users of this solution include business analysts and sales team members, who would benefit from the automated generation of client-specific newsletters. The system was designed to be scalable, enabling the client to expand the number of data sources and clients served. Ongoing maintenance and future enhancements of the solution were discussed, with plans to fine-tune the system prompts, improve scalability, and refining outputs. This proof of concept demonstrated the feasibility of using generative AI to automate newsletter creation, and with further development, the solution is expected to be rolled out more widely within the client’s organization.
Figure-01
AWS Services Used
AWS Compute Services – Lambda, Fargate
AWS Storage Services – S3
AWS Networking Services – VPC, CloudFront
AWS Management and Governance Services – CloudWatch, IAM
AWS Security, Identity, Compliance Services – IAM, Secrets Manager
AWS Application Integration – SNS
AWS Containers – Amazon Elastic Container Registry (ECR), Elastic Container Service (ECS)
AWS Database – Amazon Aurora, RDS
AWS Machine Learning – Bedrock
AWS Serverless – AWS Lambda
Third-party applications or solutions used
- Terraform Cloud
- Azure DevOps
- Serper.dev
Outcome Metrics
- Optimized operational efficiency by automating the generation of client-specific newsletters, saving approximately 60 hours of business analyst labor per week
- Enhanced the scalability and comprehensiveness of client communication processes by developing automation capable of handling data from three initial sources, 2 custom data sources plus internet search, with potential for future expansion
- Improved the accuracy and relevance of client newsletters through the use of Generative AI, ensuring tailored content for up to six sales team members
- Streamlined the information gathering process by implementing AWS Bedrock and Knowledge Base, eliminating the need for complex ETL (Extract, Transform, Load) tasks
- Reduced the time required to produce newsletters by implementing a solution that compiles targeted news for client owners, with potential for further reduction through multi-threading optimizations
Lessons Learned
- Initially, we considered using AWS Glue for complex Extract, Transform, and Load (ETL) processes. However, we learned that Amazon Bedrock and Knowledge Base could ingest HTML, CSV, text files, and PDFs directly without the need for additional ETL tasks. This streamlined the data ingestion process significantly
- We used LangChain, an external Python library, to simplify working with Large Language Models (LLMs). However, since AWS services like Bedrock are frequently updated, LangChain had not kept pace with the latest changes. In retrospect, directly using AWS’s Boto3 might have offered more flexibility and less need for workarounds
- The solution’s end-to-end process took four to five hours due to sequential LLM calls. Implementing multi-threading would allow us to make parallel calls to the LLM, accelerating the newsletter generation process
- AWS Knowledge Base proved to be effective for retrieval-augmented generation tasks. It simplified querying data and running prompts against it. This experience suggests that Knowledge Base could be a valuable tool for similar use cases with future clients
- It was noted that the recommendation to use AWS Knowledge Base came from our team rather than the client’s initial request. This proved to be a successful approach, demonstrating the importance of staying abreast of AWS offerings and confidently recommending the most effective solutions
- Although AWS Lambda was included in the solution, its use was minimal and focused on generating metadata files for Knowledge Base and setting up the initial database schema
Summary
Vertical Relevance partnered with the client to address a significant challenge in their Sales enablement workflow. The client was seeking to streamline the process of providing actionable insights for its Institutional sales team to help them pursue new business opportunities and take proactive steps to prevent client attrition. By implementing a solution leveraging Amazon Web Services (AWS) and generative AI, Vertical Relevance automated the collection and synthesis of data, saving the client approximately 60 hours per week of analyst time. The bespoke system integrated multiple AWS services, including Amazon S3, AWS CloudWatch, AWS Lambda, and Amazon RDS, along with AWS Bedrock for AI capabilities. The resulting automated process not only provided targeted newsletters for each account owner but also proved the viability of using AWS’s generative AI to effectively create customized content. This successful proof of concept holds the potential for scalability and further refinement, setting the stage for broader deployment across the client’s business operations, thereby enhancing efficiency and allowing business analysts to focus on more strategic tasks.