Nitesh Shroff Is Teaching Hospital AI to Show Its Work

The bill arrived after an unremarkable four-hour emergency-room visit. The number was remarkable: $19,000. Preeti Bhargava, who would co-found Arintra with Nitesh Shroff, went line by line and found coding errors. The bill fell to $11,000; insurance covered $7,000, leaving $4,000 out of pocket. The episode was expensive, tedious and oddly clarifying. Somewhere between the care delivered and the amount requested, a translation had gone wrong.

Shroff and Bhargava were both PhDs in artificial intelligence. Their instinct was to inspect the system behind the surprise. Medical coding converts the language of a clinical chart into the standardized codes that determine a claim. Hospitals cannot skip the step. Yet it mixes sprawling rulebooks, local practices, detailed documentation and judgment under pressure. It was the sort of problem an engineer can spend a long evening thinking about and then, inconveniently, several years building around.

Arintra emerged from that inquiry in 2019. Its early product work included patient intake and clinical documentation, and the company eventually concentrated on autonomous medical coding. The choice gave Shroff a particularly consequential place to apply a career spent teaching machines to interpret the world. A camera has to decide what a scene contains. An autonomous car has to decide what the scene means. A coding system has to decide what a chart supports, then preserve enough evidence for a human to check the answer.

First, teach the machine to see

Shroff's route to hospital administration began far from it. He studied electrical and electronics engineering at IIT Madras, then went to the University of Maryland for doctoral work in machine learning and computer vision. His research ranged across video summarization, moving-scene recognition, depth reconstruction and robotics. In 2012, Maryland's electrical and computer engineering department named him a Distinguished Dissertation Fellow, an award for students in the final stages of research.

The subjects sound academic until you notice the repeated practical question: how can a machine extract the important structure from a flood of imperfect information? At Qualcomm, Shroff worked on computational imaging features shipped on phones. At Light, he joined when the computational-camera startup had eight employees, watched it grow to about 140 and helped ship the L16, a camera that combined information from multiple lenses. The device resembled a tiny committee of eyes trying to agree on one picture.

In 2017 he moved to Zoox. There, deep-learning models had to recognize pedestrians, stop signs and movement in real time. Safety was not a slide near the end of the presentation. It shaped the engineering. Years later, Shroff would use a blunt sentence to describe the same standard in healthcare revenue software: “You cannot afford to hallucinate here.”

“You cannot afford to hallucinate here.”Nitesh Shroff on the accuracy bar for healthcare AI

The claim is a translation

A clinical note is written to describe care. A claim is assembled to request payment. The code between them must be accurate enough for compliance, complete enough to capture the work, and fast enough not to slow cash flow. When any part fails, the consequences appear as rework, a denial, or revenue that quietly never arrives.

Shroff's shorthand for the whole journey is pleasingly spare: “Document, charge, get paid.” Documentation should reflect what happened. Coding should translate it accurately. Denial prevention and appeals should help the resulting claim survive scrutiny. Arintra began in the middle, with coding, because it is the workflow every claim must cross. The company has since expanded toward documentation intelligence, DRG validation, payer insights and denial appeals.

The crucial word in Arintra's pitch is not autonomous. It is explainable. The software connects a generated code to the language in the chart that supports it. That audit trail appears within the electronic health record rather than in a separate digital annex that staff must remember to visit. The design respects a basic fact of enterprise software: even a correct answer becomes cumbersome when it arrives in the wrong room.

One unintended use revealed the value of keeping the evidence attached. Hospitals began bringing Arintra's explanations into denial appeal letters. The system had been designed to justify its coding decisions for review and compliance. Customers found that the same justification could help defend a claim to a payer. A product becomes more interesting when users reuse its reasoning, not only its result.

Even the company name preserves a fossil from an earlier idea. During a conversation about the founders' first concept, a pulmonologist described it as an “Artificially Intelligent Triage Assistant.” The words contracted into ARINTRA. The product later moved from intake toward coding and revenue assurance, but the name stayed. It is a small reminder that companies rarely proceed in a straight line. The original problem can change shape while the habit of studying it survives.

The useful machinery of trust

Shroff names simplicity, empathy and trust as Arintra's principles. In practice, those soft nouns produce hard requirements. Simplicity means working inside Epic or athenahealth without asking a coding team to adopt a new choreography. Empathy means understanding that coders are not waiting to be dazzled by a language model; they need reliable output, visible rationale and a sensible place for exceptions. Trust means that a claim can be traced back to the chart.

The company reports that its platform now handles more than $5 billion in annual claim value across customers representing over $50 billion in combined net patient revenue. It also reports gains in compliant revenue capture and reductions in coding cost and coding-related denials. Those numbers describe a business case, but they also reveal Shroff's preferred proof: results inside the workflow.

That preference shapes sales as much as engineering. Arintra leads with pilots and expects the value to appear quickly enough to convert the evaluation into a longer relationship. Shroff also describes enterprise messaging as hierarchical. A coding director wants integrations, workflow details and an accuracy method. A revenue-cycle executive wants operational leverage and compliance. A CFO wants to understand margin and recovered revenue. One product has to survive all three conversations.

Shroff arrived at company building with a deep technical ledger: more than 30 patents and publications, plus internships at Texas Instruments, Mitsubishi Electric Research Laboratories and Cisco. Research rewards novelty; operating rewards repetition done well. A health system does not purchase an interesting experiment. It purchases a dependable process that can run across thousands of charts, tolerate edge cases and make its own behavior legible to specialists. His career has required a change in rhythm from discovering a method to making one boring enough to trust.

There is a neat symmetry in this. His research career taught systems to select the important information. His operating career requires him to select the important argument. The coding director, the revenue leader and the finance chief each see a different scene. The founder's job is to help them agree on one picture.

After the code

Arintra joined Y Combinator's Winter 2022 batch. It raised a $21 million Series A in August 2025, led by Peak XV Partners. A year later, on August 26, 2026, the company announced a $25 million Series B led by Define Ventures, bringing its stated total funding to $51 million. The plan is to broaden its reach across enterprise health systems, deepen specialty coverage and carry its coding intelligence into more of the revenue cycle.

The platform now spans more than 23 specialties across ambulatory, emergency, diagnostic and inpatient settings. Its customers include UC Davis Health, Rochester Regional Health, Mercyhealth, Meritus Health, Reid Health and Mercy Medical Center. Scale adds a demanding twist to Shroff's original problem. A model must handle not one neat rulebook but specialties, care settings, payer behavior and the local coding culture of each organization.

He does not speak about that complexity as an excuse for opacity. Quite the reverse. The wider the system travels, the more valuable its audit trail becomes. Shroff's earlier machines interpreted pixels and motion; the current one interprets documentation, guidelines and clinical context. In every case, confidence is earned by connecting an answer to what the machine actually saw.

The $19,000 bill is now part of company lore, a wild number tamed into an origin story. Its lasting contribution was to make an invisible workflow visible to two engineers. Shroff has spent the years since turning that moment into infrastructure: read the record, make the decision, preserve the reason. In a field eager to sound intelligent, showing the work may be the more durable ambition.