Joint blog by Behavox and Google Cloud Engineering team

Financial services companies are subject to a variety of regulations designed to protect customers and to prevent fraud and other illegal activity. However, complying with these regulations can be a challenge, especially for large financial services companies with a large number of employees. In order to protect their customers and their businesses, these companies need to be able to quickly and effectively identify and respond to threats. Google’s PaLM 2 and Google Cloud’s generative AI tools are a powerful combination that can be used to build innovative solutions for financial services compliance.

How Google Cloud’s LLM and gen AI services can help financial services compliance clients

PaLM 2 is a large language model (LLM) that can be used to process and analyze text with high accuracy. Google Cloud’s enterprise offerings for gen AI include a variety of text, chat, and embedding models, LLM-powered applications and services that can be used for a variety of tasks, including risk assessment, and identifying potential compliance violations.

By combining PaLM 2 and other Google Cloud gen AI tools, financial services companies can build innovative solutions that can help them to:

  • Automate the processing of large volumes of text data quickly and efficiently.
  • Extract key information from text, such as entities, relationships, and sentiment.
  • Identify potential risks and generate reports.
  • Respond to threats more quickly.
  • Reduce the risk of market abuse and other malicious activity.

A Case Study: Helping a financial services client build an innovative solution

Google Cloud and Behavox, a leading provider of compliance solutions for the financial services industry, recently collaborated to build an innovative solution using PaLM 2. The solution, which is based on the latest PaLM 2 model and Behavox’s proprietary domain knowledge, is designed to process millions of communications and identify potential threats or non-compliant signals with high accuracy.

Behavox was looking for a way to improve its fraud detection capabilities. Behavox wanted to integrate PaLM into its compliance solution to improve the quality of its alerts. The PaLM family of models are trained on massive datasets of text and code, enabling them to learn the nuances of language, learn more complex patterns and relationships in the data, and generate more accurate and insightful results. PaLM’s ability to identify subtle patterns in the data that would be missed by smaller models helps Behavox improve the accuracy of its solution and reduce the number of false positives.

Google Cloud and Behavox worked together to design and implement a new fraud detection solution using PaLM 2. The solution is based on the following key elements:

  • Prompt engineering and tuning: Google Cloud and Behavox worked together to design and test a variety of prompt engineering and tuning techniques to improve the accuracy of the solution.
  • Retrieval Augmented Generation (RAG): RAG is a technique used to provide contextualized industrial data directly to the LLM as specific content to use when formulating a response.
  • Google Cloud highly scalable services: LLM offerings from Google Cloud are complemented with a variety of managed services, such as cloud functions and matching engine that can help accommodate heavy workloads.
  • Behavox’s domain knowledge: The solution leverages Behavox’s proprietary, domain-specific, and high quality dataset that was reviewed and prepared by multiple compliance professionals to ensure trustworthiness.

The solution implementation also managed to comply with other business requirements, such as explainability. In a compliance environment, when flagging messages for auditing purposes, there are many governing policies and AI needs. In addition, PaLM 2 is a multilingual large language model that supports over 142 languages, which enables Behavox to support customers in multiple countries, such as Mexico, Colombia, Japan, Canada, the UK, and the United States, delivering effective compliance solutions on a global scale. Google Cloud also worked with Behavox to evaluate the effectiveness of the solution, and preliminary commercial results communicated by the Behavox team came in line with our expectations, outpacing the performance of many other LLMs available in the market by ~20% points in some cases as measured by F1 scores.

The success of the project was inevitable due in large part to the close collaboration between Google and Behavox. Engineers from both companies worked together to design, implement, and test the solution. This collaboration enabled the teams to share expertise and best practices, and to quickly identify and resolve issues.

“At Behavox, we are passionate about using AI to solve complex challenges. Google AI engineers share our passion and have worked tirelessly to help us innovate and build scalable solutions. Their expertise and collaborative work culture are invaluable and unmatched. We are very grateful for the opportunity to partner with Google,” said Erkin Adylov, Founder and CEO of Behavox.

Future Experimentations

Google Cloud LLM offerings include many capabilities such as Reinforcement Learning from Human Feedback (RLHF), a method that can boost LLM performance compared to other methods. Google Cloud is working with Behavox to introduce such advanced capability into the solution, which will help it maintain top performance over time.

Better together: Behavox + Google Cloud

In today’s fast-paced financial landscape, regulatory compliance is paramount. All financial institutions around the globe are required to audit communication channels to be in compliance with regulations. But the current process of filtering, analyzing, and categorizing thousands and millions of messages is role-based and requires manual auditing, which makes it slow, expensive, and prone to human error. The current legacy lexicon systems produce lots of false positive alerts, making the process a huge burden for those in charge. Together, Behavox and Google Cloud built an innovative solution powered by PaLM and using Behavox proprietary domain knowledge and business intelligence to automate this rigorous process, allowing financial institutions to gain 98% efficiency over existing role based systems.

Scale the solution

To build a cloud-native solution at scale on Google Cloud leveraging the Vertex AI platform, Behavox harnessed a number of cloud-managed services to architect their Google Cloud Marketplace app. Behavox solution started by ingesting reference data from various sources in Cloud Storage buckets; parsing and preprocessing the data using Cloud FunctionsDocument AIcloud computing VMs and proprietary business logic; and leveraging the PaLM embedding API and Vertex AI Matching Engine to index and use nearest-term search on the generated embeddings at scale. And for real-time API inferencing, Behavox used Google Cloud Apigee API Management to track, monitor, and authenticate incoming requests, and pass them on to auto-scaling Cloud Functions, which in turn call Vertex AI prediction services hosting Google LLM models. Behavox also used Bigtable as a NoSQL serverless database service to store the actual label text. Finally Behavox used App Engine to quickly wrap the solution with a presentation layer add-ons.