How We Built Our First AI Prototype

Lucas Chapin

Head of Data

Want the AI boost, but without the hallucinations, transparency problems, and security concerns? We do too! Follow our progress as we develop compliance-grade AI tools to support AML investigations.

Background: What Does Compliance AI Look Like?

For companies developing AI automation tools, 2023 has been a banner year. For Hummingbird, it’s meant excitedly pursuing new tool and feature innovations built on a large language models (LLM) infrastructure.

There’s a good reason for this. With huge amounts of unstructured and semi-structured data to parse, AI tools make a ton of sense for the compliance market. Imagine an intelligent assistant that can help navigate and find information across many disparate data sources during a complex AML investigation. Even better, imagine if this assistant never forgets anything!

What follows is the story of our first foray into bespoke AI tools for compliance work: a tool we’re calling “ProfileAI”.

Building a Prototype

Before releasing ProfileAI to the general public, we wanted to aid development by soliciting early customer feedback. So we built a prototype, which we unveiled at the ACAMS conference in October.

ProfileAI is designed to assist with customer diligence, particularly for complex cases such as EDD. Why Customer diligence? It like a good test of an LLM since investigators often collect tens or hundreds of pages of text about a subject across multiple data sources, and we thought we could save investigators time by providing an interface for text search and analysis.

Here’s a peek at what the prototype looks like:

At first glance, this might look familiar to other AI tools you’ve tried. There’s a large language model whose interface takes the form of a chatbot, allowing the user to ask questions or make requests in plain language.

What makes ProfileAI Different

ProfileAI is a tool built for the rigors of compliance work, meaning this isn’t your run-of-the-mill chatbot. There’s a lot more going on under the hood. In particular, our prototype is built according to the principles we believe are required for compliance-gradeAI.

For example, ProfileAI is:

  • Fully cited and sourceable
    Hallucinations are a major concern with LLMs, particularly in the compliance industry. To combat this, we utilize a multilayer architecture that separates the retrieval step from response generation. (The technical name for this is a retrieval-augmented generation, or “RAG,” architecture.) A RAG architecture allows ProfileAI to find exact text references to questions directly source documents, and to then display those text references in the UI (along with a link to the source document for further reading). Not only does this prevent hallucinations, but – in a must-have for compliance professionals – the model shows its work No more guessing where the answer’s coming from!

  • Auditable
    Remember the old compliance adage: if you didn’t document it, it didn’t happen. The ProfileAI prototype includes: 1.) a copy/paste button for adding results to your case; and 2.) a download feature to save the chat history to the current review.

  • Secure
    All of our AI models are hosted internally within our Amazon VPC. This means no data is ever sent to a third party server, and you can ask ProfileAI questions without fear of leaking sensitive data. We also have rigorous internal isolation standards that prevent data exfiltration.

  • Easy to use
    We get it – staring at a blank search box can be intimidating. But don’t worry. While ProfileAI allows you to conduct an open search, it also provides prompts to reduce the cognitive overhead. As an added bonus, this also enables faster response times as ProfileAI is able to precompute answers as soon as documents are uploaded. In the future, we intend to support customization of these prompts which would fit into our design philosophy of heavily customizable workflows.

Feedback and Next Steps

We’re happy to report that customer reaction to the ProfileAI prototype has been a resounding, “This is great!”

Actually, the excitement for Profile has been so strong that that reaction is typically followed by, “Do you think it could also do X, Y, and Z?” In almost every case, the answer is yes...soon. We’re focused on building AI compliance products that are fast-moving and using the latest technology, but (as we’ve already mentioned) not at the expense of safety, security, or industry-readiness.

Wondering what the future of ProfileAI holds? Some of the ideas we’re currently working on include:

  • Inclusion of additional data sources
    We’re seeking to include open source or partner integration tools that will allow ProfileAI to go even deeper with its answers. For example, is there any adverse media concerning the beneficial owners of the business? Does the business’s website look legitimate? Are there any discrepancies between the products or services offered on the website and the business’s transactions?

  • Even deeper domain knowledge
    Imagine asking questions like Is there anything that stands out in these documents? or What information is missing? These questions would be far too vague for an off-the-shelf LLM model. With more fine tuning of our models, additional prompt engineering, and orchestration within a broader software product, however, we can give ProfileAI a high level of contextual awareness where AML investigations are concerned.

  • More fluid product integrations
    Chat interfaces are popular right now (in part because they’re so widely available), but we believe the future of LLM product design will involve seamless, multimodal product interfaces rather than standalone chatbots. We intend have AI embedded frictionlessly within other parts of our platform (such as automations). And as always, we’ll do so while maintaining our principles around security and auditability.

Moving Forward

Right now, we’re focused on bringing ProfileAI (as well as several other AI projects) to general access. As we do so, we continue to discover new ways for our compliance-specific LLM features to be useful. For example, we’re investigating expanding ProfileAI beyond customer diligence, transforming it into a semantic search tool applicable across a wide variety of investigative workflows and data sources. While we’re not yet ready to go into too much detail of exactly what this will look like, it’s fair to say that ProfileAI is shaping up to be one of our most exciting features to date.

Stay Connected

Subscribe to receive new content from Hummingbird