Goodbye, Ctrl+F: AI startup aims to help financial analysts sift through business documents with chatbot tool

A new startup wants to make financial analysts more efficient with its chatbot tool that uses artificial intelligence to search publicly available business documents and answer research questions in natural language.

Finpilot is led by CEO Lakshay Chauhan, a longtime machine learning engineer at hedge fund Euclidean in Seattle. He is joined by John Alberg, the co-founder of the fund, who will play a close advisory role. Both Chauhan and Alberg are founding members of the board.

The startup recently secured $500,000 in a convertible stock from Madrona Venture Labs and Ascend, both of which are investing heavily in AI companies.

We were drawn to Lakshay and John’s deep understanding of the problems facing financial analysts and the clarity of their product solution, which enables world-class research and reporting in a fraction of the time as existing tools, he told GeekWire. Madrona Venture Labs CEO Mike Fridgen. .

Alberg is the son of the late Tom Alberg, co-founder of the Madrona Venture Group and leader of Seattle’s tech community, who died last year. Finpilot will operate independently of Euclidean, which uses machine learning to select stocks and recently launched an exchange-traded fund (ETF).

The startup is developing a ChatGPT-like interface. The tool uses a large language model (LLM) to sift through publicly available financial data from sources such as Securities & Exchange Commission filings, earnings transcripts, and investor presentations.

This addresses a common pain point in the information-gathering process, where the traditional method has been relying on the Ctrl+F keyboard function to search for documents, Alberg said. Using Finpilot, research and reporting projects that would otherwise take days or weeks could be done in hours, he said.

Finpilot allows analysts to write queries in natural language. For example, a user might ask the chatbot about the increase in comparable sales for Walmart over the past 12 months. The tool would then produce a response with links to the original source documents. Analysts can continue conversations with follow-up questions, gleaning deeper and deeper insights.

Alberg said Finpilot differs from ChatGPT in that users can interact with the data, view it and verify sources. The entrepreneur and hedge fund manager has long been a proponent of using machine learning as an investment tool.

Finpilots technology is based on an in-house LLM specifically designed for finance. It also uses some LLM-based APIs and a web app stack for app deployment. The platform is under development and the plan is to sell it under a software-as-a-service (SaaS) license. The company is actively testing it with more customers, Alberg said.

Companies like Factset, Bloomberg, and Alpha Sense offer financial data tools. SEC’s EDGAR advanced search tool allows users to search for electronic documents dating back to 2001 using keywords and names. Still, Finpilot helps companies analyze, write reports and visualize data through natural language, Alberg said.

Madrona Venture Labs and Ascend invested in Finpilot through a SAFE note, or Simple Agreement for Future Equity.

The launch of the startups coincides with a rush of Wall Street firms investing heavily in AI tools and expertise. Banks are betting that recent advances in technology can simplify data processing and risk modeling, gaining an edge as the analysis becomes ever more complex.

However, experts warn that the widespread implementation of AI in finance could have several potential downsides. One concern is the potential for AI chatbots to provide inaccurate information.

Traders, who have a fiduciary duty to base investment decisions on reliable data, may face legal uncertainties when using artificial intelligence. Tracking chatbot training data can be challenging, hampering the ability to identify the source behind a fake data point used to make an investment decision.

This will be a challenge for Finpilot, but the company aims to provide users with the highest accuracy, speed and reliability, Alberg said.

Large language models are known to build information and are too slow for many tasks, he said. We have developed many techniques to deal with these problems.