Though Dr. Sanjay Rajagopalan is neck-deep in generative AI as the chief design and strategy officer at Vianai Systems, he’s an engineer by training—beginning his career at Caterpillar, building transmission systems for large earth-moving equipment.
Though that work and finserv AI development may seem far apart, Rajagopalan believes innovations like ChatGPT lowered the barrier for anyone to experience AI’s power, provided they’re willing to try.
“The best way as a product person to leverage that newfound power is to get your hands dirty,” he said. “The least you can do is jump into it and try it out, experiment with it and try to do it quickly to get it to work in your particular area.”
Vianai’s latest excavation is hila, touting it as the “first AI-powered chat assistant” for financial professionals. Vianai is providing free access in its research phase, with nearly 2,000 users in “various financial professions,” according to a spokesperson.
Dr. Vishal Sikka founded Vianai Systems in 2019, and it currently includes about 110 employees spread across Europe, Israel and the U.S. Sikka was a former board executive at SAP, where he met Rajagopalan. Rajagopalan’s relationship with senior Vianai members goes back decades, with a core group working together continuously in different companies and contexts.
Though Vianai’s released products since its inception, hila is the first released since the “tsunami” of ChatGPT hit the world, and while the team reimagined its old products to incorporate the new tech, they wanted to go further. A four-person team progressed from an “on-paper” idea for hila to a working prototype in weeks, according to Rajagopalan.
“We wanted to start something completely from scratch and reimagine it in this new world, rather than retrofit this into something that had been built in the old world,” he said.
One benefit of using hila compared to ChatGPT is that the latter is trained on information before November 2022, but advisors need more up-to-date information from earnings calls, news and regulatory documents. (Though many companies are now clued into this shortfall, Rajagopalan said Vianai was at the forefront of addressing it.)
While many are wowed by generative AI’s ability, large language systems really just predict the next word, and with such a large corpus of information, the most likely world also tends to make sense, according to Rajagopalan. But when this doesn’t work, it can fail in big ways, which are known as “hallucinations.”
“Sometimes, it would go off the rails, and it would go off the rails with extreme confidence,” Rajagopalan said. “People would ask it a question … and it would give an extremely confident answer that was completely wrong. You now have a piece of information which you may have to act on to make your investment decision. This is obviously unacceptable.”
hila reduces the number of hallucinations to “near-zero,” according to Rajagopalan. When there’s only one correct answer, hila’s model simply generates a database query, allowing users to get specific answers using natural language.
The tool also allows users to upload a document and use hila to generate an answer from data included in it in natural language. For example, one can ask hila to summarize key points in a lengthy report, which isn’t possible with a keyword search.
“The only way you could do that before was to ask another person,” Rajagopalan said. “Now, you can ask the system to summarize the document for you, and it can do a pretty good job.”