An Accenture-backed tech company that helped NASA launch the Artemis moon program has unveiled what it claims to be Wall Street’s first LLM for wealth advisors that does not hallucinate.
The Stardog Voicebox Wealth Assistant avoids hallucinations because it doesn’t use LLMs to retrieve information directly and then give it to the user, according to Stardog founder and CEO Kendall Clark.
He and Matt Lucas, a former Morgan Stanley executive who joined recently as the field CTO for financial services, spoke to The AI Innovator about how they built Voicebox and why in the future they plan to offer a no-hallucination guarantee to clients.
The AI Innovator: First of all, what a cool name you have! Tell me how the name Stardog came to be.
Kendall Clark: When we started developing the first version of the product, one of our founders, Mike Grove, asked me to watch some critical first automated tests passing, and he was listening to a song called ‘Stardog Champion.’
I hated the song but loved the name, and we decided on the spot to call the product – and then the whole company – Stardog, since everyone loves dogs and data integration ‘star schemas’ are an important part of how Stardog creates an enterprise semantic layer.
I understand you debuted what you say is a 100% hallucination-free AI assistant for wealth advisors. Tell me first about the product. How does it work? Who uses it?
KC: Stardog Voicebox Wealth Assistant pulls together far-flung research and reference data for wealth advisors and use gen AI to provide instant access to insights so that they can provide real-time advisory services to clients. The value is time-to-insight and avoiding the timely and costly back-office data science and research function to address client questions and needs. We can’t yet talk about our first few customers, but it’s in use and adoption is growing.
How did you manage to solve the problem of hallucinations? Please be specific.
KC: All LLMs hallucinate, and that’s a big problem for regulated industries. We solved this problem in Stardog Voicebox by using LLMs differently than other AI firms; rather than using LLM as a source of ‘answers,’ we use it as an algorithmic detector of human intent. In other words, we use it to turn plain language into knowledge graph queries into our platform, which has in turn connected to enterprise data sources and silos.
When our competitors’ LLMs hallucinate, the user sees that misleading output directly and has to be very careful about using that material. When our LLM hallucinates, the customer never sees that misleading material but rather — about 5% of the time — sees instead a message that the users’ question cannot be answered. While no one loves having to re-phrase a question, we think it’s a much better user experience than always giving an answer that sometimes contains misleading hallucinations.
Our approach is based on Semantic Parsing rather than RAG and that’s a key insight that we had – to avoid hallucinations from misleading users, you cannot allow the LLM to speak to the user last. In Stardog Voicebox, the users’ data speaks last and that’s a more trustworthy experience.
How do you balance the need for accuracy and creativity of Voicebox?
KC: By listening carefully to our customers who are pursuing high stakes use cases in regulated industries, we understood quickly that this balance should be tipped entirely in favor of accuracy. The question gets the tradeoff exactly right: Voicebox isn’t very creative because the requirement is accuracy.
For a different use case, for example, an LLM-powered life coach or creative writing assistant, the tradeoff should probably go the other way. But we focus on financial services, life sciences, manufacturing, and national security markets, and it’s clear in those areas that accuracy is more important than creativity.
At the same time, Stardog Voicebox can use foundational LLMs directly for those cases where our user can intentionally choose to favor creativity. We support both and are careful to help the user understand which Voicebox responses are based on enterprise data and which may contain hallucinations.
Do you provide any guarantee that Voicebox will not hallucinate?
KC: We don’t currently guarantee no hallucinations contractually, but we do anticipate providing that guarantee as market conditions shift. We think the market is still mostly in the PoC (proof of concept) phase where the guarantee is less significant, but we also see many early signs of market maturation and a broad shift from experimentation to real adoption and will react accordingly.
Give me a sample use case of how a wealth advisor can use Voicebox. Please be detailed.
Matt Lucas: One of the biggest challenges advisors have today is to give their clients access to timely, deep insights in real time. When clients ask challenging questions, this often triggers internal data investigations, navigating through complex prospectus documents, and could jeopardize a potential deal. Stardog Voicebox, by connecting several data sources inside the firm, is able to immediately respond with fast and accurate information for questions that are challenging to answer quickly today. We help put deep institutional knowledge into a fast, accessible format, enhancing the value that advisors can provide to prospects and clients.
Which clients are using Voicebox and how are they using it?
ML: While we can’t disclose the specific financial institutions where our technology is being deployed outside of what we list on our website, we are proud to support several G-SIFI (Global Systemically Important Financial Institution) companies with our technology, helping some of the most complex financial institutions in the world enhance the value they bring to clients. Outside of financial services, we also have enabled many global 2000 organizations in consumer goods, life sciences, energy and manufacturing, and government entities with our technology.
Is it an enterprise-only product or is it also for individual financial advisors?
KC: Our commercial focus is on the enterprise market and the platform includes all the required capabilities and compliance ratings necessary to operate in the world’s most heavily regulated businesses. While we don’t discourage individual financial advisors from engaging, we’re focused on the enterprise market.
What’s next for Voicebox?
KC: We’ve been very focused on performance, that is, answering 95% of all questions in under 2 seconds, and have made great progress toward that goal. We’re also expanding Voicebox from financial services to core use cases in manufacturing and life sciences, primarily around end-to-end visibility of global supply chains so that organizations can improve quality of service, avoid supply chain shocks and outages, and drive down costs associated with manufacturing, logistics, ESG, and product management.
What’s next for Stardog?
KC: We’re excited about national security opportunities in both defense and intelligence around supporting decision-advantage and mission-focus in the U.S. and among the Five Eyes partner nations (an intelligence-sharing alliance among five English-speaking countries: U.S., U.K., Canada, Australia, New Zealand).
On the product side, we’re adding support for unstructured data, such as enterprise documents, to our existing support for structured data (or database records), as gen AI-powered insight for knowledge workers requires comprehensive enterprise data access. We’re excited to be the first or one of the first gen AI vendors to offer an AI Data Assistant that is hallucination-free and comprehensive with respect to all enterprise data.
Be First to Comment