Five Minutes with Stardog's Matt Lucas

INTERVIEW
Accenture-Backed Stardog: AI in Finance & Rocket Science

AI Street interviewed Matt Lucas, Field CTO for financial services at Stardog, a company using AI and knowledge graphs for data challenges in finance and other industries. Based in New York City, the company uses knowledge graph technology, which organizes information to show an interconnected map of data.

Stardog, which is backed by Accenture's investing arm (the consulting firm is also a client), employs generative AI to build these knowledge graphs, aiming to make AI more reliable and efficient, particularly in addressing the issue of AI hallucinations. The company says its own LLM, Voicebox, is fast, accurate, and 100% hallucination-free.

Lucas shares his thoughts on what kind of ROI clients can expect, what's overhyped (General purpose LLMs), underhyped (Small LLMs) and current use cases. Before joining Stardog, Lucas was most recently Executive Director of Technology at Morgan Stanley.

The following interview has been edited for clarity and length.

What is Stardog and how does it address AI hallucinations?

Stardog is a knowledge graph and AI platform provider specializing in using AI to solve complex data challenges across various industries, including finance. We focus on safe, generative AI using knowledge graphs. By structuring data in graph format, we can significantly reduce or eliminate hallucinations, providing more reliable AI solutions.

What are knowledge graphs?

Knowledge graphs are a way of structuring data that emphasizes relationships between different pieces of information. They're important for AI in finance because they help organize sprawling data, make AI adoption more efficient, and improve model accuracy and reliability. As AI in finance evolves, knowledge graphs are becoming a crucial foundation for advanced, trustworthy AI systems.

How does Stardog's method differ from traditional Retrieval-Augmented Generation (RAG)?

RAG, or Retrieval-Augmented Generation, feeds documents to an LLM (Large Language Model), which isn't foolproof. We take a different route:

1. Put all firm data into a graph form
2. Use LLMs to navigate this structure
3. Convert user questions into precise queries
4. Provide traceable, accurate responses

This approach works with various data types - not just documents, but databases, data lakes, and complex data structures.

Can you give an example of how Stardog's technology is used in finance?

Imagine a wealth advisor with a high-net-worth client requesting a product with specific criteria - say, less than 20% tech exposure and limited exposure to certain businesses. Traditionally, this would require extensive research. With our system, the advisor types this question and instantly receives a response like: "Here are three funds meeting your criteria. Fund A is 30% tech, meets other requirements. Here's the prospectus link." It's like having an economist at your fingertips, dramatically cutting research time.

What are some other use cases? 

We focus on several critical areas:

1. Anti-Money Laundering (AML) and Know Your Customer (KYC): Quickly identifying suspicious patterns across complex networks of transactions and relationships.
2. Fraud detection: For instance, identifying potentially risky transactions by uncovering multi-step connections to nefarious entities.
3. Risk management (both front and back office): Providing comprehensive views of risk exposure across various financial instruments and counterparties.
4. Wealth advisory: As mentioned earlier, enabling rapid, personalized product recommendations based on complex criteria.

Essentially, we excel anywhere data is sprawling and hard to navigate.

How does Stardog's solution compare cost-wise to developing custom AI models?

Our solution is often more cost-effective, especially for medium-sized firms. Instead of investing millions in training and maintaining custom LLMs, companies can leverage our pre-trained models and knowledge graph technology. This approach allows firms to achieve similar or better results at a fraction of the cost and time investment of developing custom AI solutions.

What ROI can clients expect, and what does the implementation process look like?

We aim for a significant ROI, typically 300% or higher, so for every dollar invested in our solution, clients can expect to see at least three dollars in return through increased efficiency, better decision-making, and reduced operational costs. Implementation is relatively quick - usually an 8-10 week cycle. This process involves connecting our system to the client's data sources, applying our pre-defined industry models, and fine-tuning the system for the client's specific needs. This is a major advantage; instead of spending 18 months training an LLM, clients can start getting valuable insights in about two months.

How does this fit with compliance?

Our approach provides full traceability, which is crucial for regulatory compliance. We can show exactly how we generated a response, from the initial query to the data sources used. This transparency is vital when regulators ask how certain information or decisions were produced, allowing firms to demonstrate their compliance clearly and efficiently.

Can you share any notable clients or use cases outside of finance?

While we can't disclose all our clients, one notable example is NASA. They use Stardog to help manage the relationships between parts within the rocketry they're building. 

What LLM does Stardog use? 

Currently we’re using an open-source version of Llama, but we're not tied to any single model. Our strength is in applying these models to specific use cases like query conversion, predictive modeling, and anomaly detection. We adapt our approach based on what works best for each task.

What's overhyped and what's underhyped in AI?

Overhyped: General-purpose LLMs. They're becoming commoditized, with many companies offering wrapper solutions on models like ChatGPT. While valuable, their use cases are limited due to hallucinations and deployment costs.

Underhyped: AI agents and small language models for specific tasks. This targeted approach is powerful and efficient. We're seeing growing interest in applying AI to specific business problems rather than trying to solve everything with one large model.

What's next for Stardog in the financial sector?

We're expanding our ready-to-go solutions, with a focus on specific verticals. For example, we're currently offering a financial wealth advisor solution to the market called Stardog Voicebox Wealth Assistant. This solution aims to streamline the advisory process, making it faster and more accurate.

What's been most surprising?

The eagerness to adopt AI coupled with a lack of deep understanding. Many firms, after being burned by overhyped solutions, are now seeking more practical, explainable AI. There's a growing appetite for solutions that solve specific business problems and provide clear ROI.

Thanks for reading!

Drop me a line if you have story ideas, research, or upcoming conferences to share. [email protected]