At the cutting edge of applying generative AI to the world of finance and investing is Bridgewater Associates, a pioneering investment management firm. They have developed what they call the “Artificial Investment Associate” or AIA — a platform that integrates large language models and other AI capabilities to assist their investment research process.

Recently, Swami Sivasubramanian, senior vice president of AI and ML at AWS, sat down with Aaron Linsky, CTO of AIA Labs at Bridgewater, about how they are leveraging models like Claude from Anthropic’s AI research group via Amazon Bedrock. Bedrock provides an abstraction layer that allows Bridgewater to plugin the latest LLMs as they become available.

“Bedrock is a great abstraction layer over that,” said Linsky. “There’s other benefits that I’m happy to go into as well. But first and foremost, what’s most important to us is that we’re getting the best models available to AIA and that we’re able to plug those in pretty seamlessly.”

One key learning for Bridgewater has been the importance of involving their expert investors and analysts directly in the process of prompting and interacting with the LLMs. As Linsky explained: “Likely the best prompt engineers for you are your subject matter experts, so your end users.”

The AIA platform allows these domain experts to construct multi-step “blueprints” that chain together a sequence of LLM queries and other data processing steps to tackle complex investment analysis questions that go beyond what an LLM could handle in a single shot.

“A blueprint is a set of steps, which is able to answer questions that might be more complex than just a single prompt question response,” said Linsky. “The agents can call the blueprints, no problem.”

Bedrock’s ability to integrate different tools has been valuable as well, such as using Amazon Textract to extract text and tables from PDF financial reports to fuel the AI’s knowledge base.

“Textract was really impressive at being able to pull out that data and get it into a nice Markdown form,” noted Linsky. “Textract is now our preferred PDF parser and we’re finding better and more relevant and accurate results in our RAG pipelines as a result.”

While still in its early stages, Bridgewater’s AIA project shows the potential for generative AI to intelligently augment and scale the work of human investment experts and analysts. Their experiences pioneering this integration of LLMs can serve as a model for other finance industry practitioners looking to harness this powerful AI technology.

Featured image: Credit: AWS