Safe Financial AI with Claude 2 and AWS Bedrock
Generative AI is incredibly exciting, and prompt engineering is surprisingly accessible, but applying these technologies while protecting consumer privacy is really, really hard. Doing all this in the heavily regulated financial industry is even harder. M1 operates mainly in two financial verticals: investing (stocks, bonds, etc.) and banking (savings and personal loans) - and each of these areas has their own regulations around privacy, communications, compliance, etc. that present particular challenges for building new AI-powered applications and features.
In this talk, I’ll propose a solution that can satisfy these requirements, using a privately hosted Claude 2 model on Amazon Bedrock, with full control over data retention, access and encryption. With this solid foundation, we are then free to cherry-pick prompt engineering techniques from research, and make a few small tweaks to comply with real-world regulatory and business requirements.
In particular, we’ll look at financial time-series data. LLM’s are a lot more capable at reasoning about numeric and tabular data than one might expect, and I’ll discuss a few research papers on what they can and can’t do. But regulators impose other limitations, especially around investment advice, and I’ll also talk about safety mechanisms and other controls that can help us ensure that our AI products are following all applicable laws and regulations.
