A solo AI project from a Columbia grad is streamlining daily operations at a multi-billion-dollar wealth firm by automating critical, time-consuming tasks.
When Vipul Harihar (MS '23) was hired by GenTrust, a wealth management firm with offices in New York, Miami and Puerto Rico, the expectations were straightforward: clean up code, support internal systems, and help modernize a traditional trade operation.
But when company leadership asked if he could "do something with AI," he took it as a serious engineering challenge. Four months later, he had designed and built an AI-powered proprietary operating system that integrates with internal and third-party software, which is now the core of the firm's trading technology platform.
"They brought me in to clean up legacy systems," Harihar says. "What started as maintenance turned into a complete reinvention."
The platform, launched in July 2025, uses the same kind of language model that is behind tools like ChatGPT, but adapted into a secure system built for financial operations.
Many AI tools are unusable in wealth management given their tendency to hallucinate, and the risk that user data could be shared or used to train models. Drawing on his education at the Data Science Institute, where he learned both the theory behind AI and how to apply it in high-stakes, real-world settings, Harihar developed a solution that worked around both issues. Harihar built a custom system that transforms thousands of internal documents, workflows, and compliance rules into information the model can draw on and use. Crucially, the system doesn't share any data.
He also reduced hallucinations by limiting what the model is asked to do. The platform operates within tight boundaries, using clear instructions and limited options to keep its answers accurate.
Unlike most language models—which work by generating the most statistically likely answers without grasp of meaning—this system asks clarifying questions when it's unsure.
Beyond automation and workflow coordination, Harihar focused heavily on user experience, and the range of staff, from traders to administrative staff – who would have to use it. The multilingual system adjusts tone, language, and detail based on the user.
"If someone isn't a statistics person, the answer has to make sense to them too," Harihar says. "I wanted to make something people actually want to use."
Perhaps most remarkable is that Harihar built the entire system himself—front end, back end, mobile integration, dashboards, data pipelines, and even the product's visual design. The total development budget was $50 a month, covering access to the language model API. With GenTrust fully onboarded, Harihar and firm leadership are now exploring ways to bring the system to other firms.
"I didn't come to DSI with all the answers," he says. "But I left with the ability to build something from start to finish. Columbia teaches you how to operate without a fixed blueprint."
