
Share
Thin-file borrowers aren't high risk, they're unmeasured. Data scientist Pragya Kumar argues the industry has been confusing absent data with absent creditworthiness, and her patented anomaly-detection work suggests a fix.
Most lending decisions in the US today never touch a human. A model scores you first, and if that model can't parse your file, the application often just dies quietly. No denial letter, no appeal, no explanation. Just silence.
That's the failure mode Pragya Kumar has spent her career poking at. She's a data science leader and named inventor on multiple US patents, and she builds the machine learning systems that sit behind small business lending and insurance underwriting. She doesn't frame this as advocacy work. She frames it as an unsolved measurement problem.
The numbers back up how big that problem is. The FDIC's most recent national survey found 4.2% of US households are unbanked, with another 14.2%, 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, and close to 10% of adults have a file too thin or stale to score at all.
It gets worse on the small business side. The Federal Reserve's latest Small Business Credit Survey found only 42% of employer firms that applied for financing got the full amount they asked for. That's despite small businesses making up 99.9% of US firms and driving nearly 89% of net job growth in the most recent year the SBA measured. LIMRA's 2025 research adds one more data point: roughly 100 million Americans lack adequate life insurance coverage.
"If the model cannot see you, the system cannot serve you," Kumar says. "A very large number of Americans are, in a statistical sense, invisible."
Kumar didn't start in finance. Early in her career she built sensor networks and predictive models for grain storage facilities, the kind of system designed to flag spoilage before anyone walking the floor could see it. "The grain looked fine," she recalls. "The model knew otherwise."
That work generated four granted US patents. Two of them, US Patent 11,307,062 and its continuation US Patent 11,592,322, cover machine learning methods for remote monitoring and predictive quality management. The other two, US Patent 12,013,257 and US Patent 12,069,770, extend the same anomaly-detection logic to a self-moving sensor pod and a multi-sensor probe that tracks temperature and gas levels inside stored crops.
The underlying pattern is simple: surface-level signals lag reality, and a well-built model should catch the gap before a human notices anything wrong. Years later, sitting in a meeting about loan applicants the system couldn't score, Kumar recognized the exact same structure. "These people are perfectly creditworthy," she remembers thinking, "and the system simply cannot see it yet. Same problem. Different silo."

That reframe isn't just a nice turn of phrase, it's a critique of how most credit models get built. Traditional scoring systems are trained almost entirely on people who already had access to credit. That makes them very good at recognizing the already-included and nearly blind to everyone else, because the training data never contained those people in the first place. It's a textbook dataset bias problem, just one that shows up as a closed account instead of a misclassified image.
"We are not declining risk here," Kumar says of how thin files typically get handled. "We are declining to do the work of measurement."
That distinction matters for anyone who builds or audits these systems. A model that rejects a thin file isn't necessarily making a risk judgment. More often, it's hitting a wall where its feature set runs out, and the easiest default is to treat missing data as disqualifying rather than as a prompt to look harder. Kumar's argument, shaped by years of watching anomaly-detection models dig signal out of sparse sensor readings, is that the industry has been choosing the lazy default far too often.
Kumar holds a Master of Engineering in Industrial Engineering and Operations Research from UC Berkeley. She now leads data science for small business and insurance product lines, applying the same instinct from her grain-monitoring patents: when data is sparse, push the model to extract more signal before it defaults to rejection. That's a harder engineering problem than it sounds. Sparse, noisy, or asymmetric data sets are exactly where models tend to either overfit on proxies or fall back on crude heuristics, and getting it right requires the kind of careful feature engineering and anomaly-detection thinking Kumar built her early patents around.
Kumar frames closing the small business financing gap as straightforward arithmetic. Get more creditworthy applicants correctly scored, and you unlock job creation and economic growth on a national scale, not a marginal one. Given that small businesses already account for nearly 89% of net job growth, even modest improvements in scoring accuracy could move that number meaningfully.
The core issue here isn't exotic. It's the same dataset bias problem that shows up across computer vision and NLP: a model is only as representative as the population it was trained on, and when an entire demographic is systematically underrepresented in training data, the model doesn't fail loudly, it just fails silently, by never flagging them as worth a closer look.
What's notable about Kumar's framing is that it treats "I don't have enough data on this person" as a solvable engineering challenge rather than an acceptable business outcome. Her background in anomaly detection for physical sensor systems, where the whole point is catching problems invisible to a human observer, maps surprisingly well onto credit scoring, where the problem is catching creditworthiness invisible to a model trained on the wrong population.
Whether the lending and insurance industries actually adopt that mindset at scale is a separate question from whether the technique works. But the numbers, 7 million adults with no credit file, 19 million underbanked households, 100 million Americans underinsured, suggest the addressable market for "see more, reject less" modeling is not small. For an industry built on being right about risk, failing to even attempt the measurement looks less like caution and more like a gap in the engineering.
Tags
Original Sources
The invisible majority: How one data scientist is teaching machines to see the Americans finance forgot
↗ https://venturebeat.com/business/the-invisible-majority-how-one-data-scientist-is-teaching-machines-to-see-the-americans-finance-forgot
About the author
Kai built ML infrastructure at a Bay Area startup before developing an obsession with transformer architectures and inference optimisation that eventually pulled him out of product work entirely. A stint at a compute research lab sharpened his instinct for what actually matters in a model release versus what is marketing. He writes from the inside — from the perspective of someone who has debugged the systems he is describing at three in the morning. He is allergic to hype and instinctively drawn to the unglamorous plumbing questions that everyone else skips over.
More from The Engineer →This Week's Edition
18 November 2026
29 articles
Related Articles

National Academies Report Calls for Faster, Tougher Retraction Standards
Models & Research · 5 min

FAMU and Sandia National Labs Sign New Five-Year Research Agreement Spanning AI and Quantum Computing
Models & Research · 5 min

Nobody Is Tracking AI Hallucinations in Customer Service, New Survey Finds
Models & Research · 6 min
Related Articles

National Academies Report Calls for Faster, Tougher Retraction Standards
Models & Research · 5 min

FAMU and Sandia National Labs Sign New Five-Year Research Agreement Spanning AI and Quantum Computing
Models & Research · 5 min

Nobody Is Tracking AI Hallucinations in Customer Service, New Survey Finds
Models & Research · 6 min
More Stories
© 2026 Cedar & Bloom. All rights reserved.