Pragya Kumar's patented machine learning methods are teaching financial models to read the Americans they were built to overlook
In modern finance, an algorithm typically makes the first decision about you, well before any loan officer reviews your file. When that model cannot read your file, the system tends to close quietly, sometimes with no formal answer ever given.
That distinction sits at the center of Pragya Kumar's work. A data science leader and named inventor on multiple U.S. patents, Kumar builds machine learning systems behind small business lending and insurance access, work she frames less as a mission than a math problem the industry has left unsolved.
A country of unreadable files
The scale of that problem is not small. The Federal Deposit Insurance Corporation’s (FDIC) most recent national survey found 4.2 percent of U.S. households unbanked, with another 14.2 percent, roughly 19 million households, underbanked. The Consumer Financial Protection Bureau's 2025 correction puts the number of adults with no credit record at about 7 million, with nearly 10 percent of adults holding a file too thin or stale to score.
On the small business side, the Federal Reserve's most recent Small Business Credit Survey found that only 42 percent of employer firms that applied for financing received the full amount requested, even though small businesses make up 99.9 percent of U.S. firms and accounted for nearly 89 percent of net job growth in the most recent year measured, according to the SBA's 2025 profile. LIMRA's 2025 research adds another dimension, estimating that roughly 100 million Americans lack adequate life insurance coverage.
"If the model cannot see you, the system cannot serve you," Kumar says of the underlying failure. "A very large number of Americans are, in a statistical sense, invisible."
From grain silos to household ledgers
Her path to that observation began somewhere unrelated to finance. Early in her career, Kumar built sensor networks and predictive models for grain storage facilities, systems designed to catch spoilage before it became visible to a person standing in the room. "The grain looked fine," she recalls. "The model knew otherwise."
That work produced four granted U.S. patents, including US Patent 11,307,062 and its continuation, U.S. Patent 11,592,322, which cover machine learning methods for remote monitoring and predictive quality management, and US Patent 12,013,257 and US Patent 12,069,770, which extend the same anomaly-detection approach to a self-moving sensor pod and a multi-sensor probe that tracks conditions like temperature and gas levels inside stored crops.
Years later, during a meeting about loan applicants who couldn't be scored, the same realization returned in a different form. "These people are perfectly creditworthy," she remembers thinking, "and the system simply cannot see it yet. Same problem. Different silo."
Measurement, not risk
“If the model cannot see you, the system cannot serve you. A very large number of Americans are, in a statistical sense, invisible.” — Pragya Kumar
That reframing, she argues, matters more than it sounds. Traditional credit models are trained almost entirely on people who already had access, which makes them skilled at recognizing the already-included and largely silent on everyone else. "We are not declining risk here," Kumar says of how thin files typically get treated. "We are declining to do the work of measurement."
Kumar holds a Master of Engineering in Industrial Engineering and Operations Research from UC Berkeley and now leads data science for small business and insurance product lines, where she applies the same instinct that shaped her patented grain-monitoring work: when data is sparse, a model should be pushed to extract more signal before it defaults to rejection.
An economic argument, not a charitable one
Kumar frames closing the small business financing gap as a matter of arithmetic: it could drive job creation and economic growth for the country. Whether that framing changes how the industry treats its blind spots will likely come down to how many more models learn to read what the data has been hiding all along.
The information provided in this article is for general informational and educational purposes only. It is not intended as legal, financial, medical, or professional advice. Readers should not rely solely on the content of this article and are encouraged to seek professional advice tailored to their specific circumstances. We disclaim any liability for any loss or damage arising directly or indirectly from the use of, or reliance on, the information presented.
VentureBeat newsroom and editorial staff were not involved in the creation of this content.
