How Kepler Built Verifiable AI for Financial Services — Vinoo Ganesh

Venu observes that current discussions around AI in finance primarily focus on increasing output (e.g., token maximization) rather than ensuring trustworthiness and verifiable results. He emphasizes that the historical bottleneck wasn't production but information gathering, which is now being automated by AI, leading to alpha decay as everyone accesses the same data sources like Tegus.

Venu draws parallels between verification processes in software development (CI/CD, code reviews), medicine (pharmacist checks prescriptions), and finance. He highlights the lack of a robust verification layer in financial services compared to these other sectors, noting that it often relies on an overworked VP.

Venu posits that financial data providers like Bloomberg and FactSet are utilized to shift culpability when errors occur. The belief is that because a third party vetted the information, any inaccuracies are shared across Wall Street, providing a degree of protection against individual accountability.

Venu argues that the current limitations of AI in finance, particularly its inability to produce verifiable work products like fairness opinions or investment memos, contribute to long working hours and potential burnout among financial professionals. He stresses the importance of AI moving beyond simple search technologies to generate outputs that can be justified and backed by reliable sources.

Kepler's technology allows for the consolidation of financial statements in mere seconds, a process that traditionally takes much longer. A key differentiator is that each number extracted and used in the consolidated statement is directly tied back to its original source document. This eliminates reliance on contractors often employed by companies like CapIQ and Deloopa, enabling the creation of numerically accurate financial models.

While currently focused on finance due to its reliance on numerical data, Kepler's technology has broader applicability. The speaker envisions a future where similar verifiable AI processes are applied to areas like legal case analysis (for firms like Harvey and Legora), ensuring no relevant precedents or citations are missed, and drug discovery, guaranteeing comprehensive review of research publications without overlooking critical compounds.

The speaker draws a parallel between the current state of verifiable AI and the pre-SSL era in e-commerce, when few people were comfortable sharing credit card information online. Just as security protocols dramatically increased trust and adoption in e-commerce, verifiable AI is poised to unlock significantly greater value and confidence across various industries by ensuring data accuracy and provenance.

The trend in the industry is shifting away from simply maximizing token consumption, which has become a form of rewarding vendors for high spending. Instead, there's a growing emphasis on optimization and cost-effectiveness, similar to how companies previously focused on demonstrating ROI for investments in platforms like Snowflake and Databricks. The focus now is on using the right tool for the job, even if it requires less computational power.