Prathamesh Juvatkar and the Long Apprenticeship of Making AI Useful

The first important machine in Prathamesh Juvatkar's public story was not intelligent. It was an electric motor. In his final year at IIT Gandhinagar, he and two collaborators designed a quicker way to model an electrical machine, balancing losses, cost and weight before validating the result with finite-element analysis. The project earned a place at an IEEE conference in the United States. The invitation was welcome; the airfare was not. Money was scarce, so Professor K. Ragavan used his professional-development allowance to help Juvatkar make the trip.

It is a small episode, but it carries most of the ingredients that would recur in his career: a difficult system made tractable, an institution still inventing itself, and another person willing to place a bet on him. Juvatkar belonged to IIT Gandhinagar's pioneer undergraduate cohort. There was little inherited campus lore because his class was busy creating it. He studied electrical engineering, added minors in computer science and mechanical engineering, and graduated in 2012 with the Institute Gold Medal for the highest academic performance in his discipline.

He later described the faculty's confidence in language that sounds less like nostalgia than a working principle: “The teachers had great faith in us. We performed things in ways that would not have been possible anywhere else but at IIT Gandhinagar.” Freedom, in this telling, was useful because somebody expected a result.

A campus without old rules

The newly minted engineer did not wander far from that experimental atmosphere. Four graduates, including Juvatkar and his classmate Sarthak Jain, formed GridAnts. It became the first startup incubated at IIT Gandhinagar. Juvatkar had already worked on FPGA programming at Marvell India and on a community seismic-network project during a summer at Caltech. GridAnts widened the field again: location services, connected devices, social software, whatever might become a product if given enough tinkering.

By 2014, the group had narrowed its attention to Cubeit, a mobile app built with Jain and Nithin Gadiparthi. Smartphones had filled with specialized apps that rarely cooperated. Cubeit tried to let a person gather content from different places, organize it and share a native experience without requiring every friend to install the same collection of software. It was an attempt to make the seams between apps disappear.

The startup raised $3 million from Accel Partners and Helion Venture Partners. In 2016, the fashion marketplace Myntra acquired the company and brought its team into the technology organization. Juvatkar spent a short spell as a tech lead on Myntra's Look Good platform, working on social feeds, collections and community interaction. One apprenticeship had ended in an acquisition. Another was about to begin.

Two founders who knew the footnotes

Founders are often advised to find a partner who complements their skills. Juvatkar offers a quieter criterion: find someone whose perspective you already understand. He and Jain met in their first year of college. By the time they began Nanonets in 2016, they had shared a campus, an incubator, a venture-backed startup, an acquisition and an employer.

“Sarthak and I have been friends for 10+ years. From the 1st year of BTech to creating, growing, and managing two companies.”Prathamesh Juvatkar

His argument is not that friends avoid disagreement. It is that history supplies the footnotes. “Many start-ups fail because of disagreements with the co-founders,” he has said. “But being friends with your co-founder eliminates most of your problems. Because you understand their perspectives and where they are coming from.” The claim can sound dangerously pleasant until one notices how long it has been tested.

Nanonets began with a broad machine-learning proposition and soon found a vast, beige-colored problem. Businesses were still paying people to open invoices, receipts and purchase orders, find the important numbers, type them into another system, and repeat. The files were inconsistent. The errors were expensive. The work was essential and remarkably resistant to the future.

This was an excellent market for a company with no need to look glamorous. Nanonets trained models to extract information from unstructured documents, then surrounded that extraction with the less photogenic machinery of integrations, approvals, validation and reconciliation. It learned that reading a document was only the first act. A customer cared about what happened next.

The architecture is allowed to lose

Juvatkar's technical account of Nanonets is striking for its lack of romance. The company initially used convolutional neural networks to examine images and identify objects. It considered graph neural networks. As transformers and multimodal systems became more accurate, the stack moved again. No architecture received tenure.

For a new customer, the team could train several model families on that customer's data and see which one performed best. “Right now, in the back end, we have multiple architectures,” Juvatkar explained in 2024. “Whenever we get a new customer, we train all of these models on the customer data and see which one gets better accuracy.” The sentence is almost aggressively sensible. In an industry that turns model names into team identities, his allegiance was to the test.

That pragmatism had a commercial twin. Juvatkar said Nanonets won business through accuracy, user experience and integrations. The three belong together. Accuracy without a usable interface is a laboratory result. Accuracy without connections to accounting and enterprise systems is an impressive dead end. By 2024, roughly half of the company's customers were in financial services, while manufacturing and other document-heavy sectors were growing. The United States produced about 40 percent of revenue and Europe another 30 to 35 percent.

The company said its user base quadrupled in the year before its Series B and had passed 10,000 customers. Accel led the $29 million round, joining existing backers Elevation Capital and Y Combinator. The financing brought total reported funding to $42 million and gave Nanonets more room for research, sales and the expensive business of improving accuracy by fractions that matter.

The glamour is the least reliable part

There is a useful tension in Juvatkar's public persona. He believes technology, paired with good timing, can alter people's lives. He also advises students not to hurry into starting companies. “It requires an extreme level of dedication,” he has warned. “You will have to do a lot of random and time-consuming tasks. It is not as cool as it is perceived.” Silicon Valley rarely puts that sentence on a hoodie, perhaps because it would not fit beside the logo.

Yet the warning is not cynicism. It is a description of craft. Juvatkar's career has moved through electric machines, seismic sensors, mobile content, social commerce, optical character recognition and enterprise agents. The objects kept changing. The method remained recognizable: study the system, build a model, expose it to reality, keep the parts that work.

The same instinct appears in his recent interest in document models that transform pages into structured Markdown and in his public questions about AI evaluation, business rules and enterprise context. As software agents are asked to do more than extract a number, reliability becomes harder to measure and more consequential. A document model can be graded against a page. An agent must also understand the surrounding process, know when to abstain and leave a trail another person can inspect.

This is less a departure from Nanonets' original problem than an enlargement of it. The invoice was never merely an image. It carried a vendor relationship, an approval threshold, a budget and an exception somebody had once written into a policy. Reading the pixels was the beginning. Understanding what the organization meant to do with them was the prize.

There is one last loop back to Gandhinagar. In 2014, Juvatkar's company returned to the young institute to recruit. He knew the juniors were capable because he remembered what it was like to be trusted before a reputation had arrived. IIT Gandhinagar later honored him and Jain for outstanding entrepreneurship. The institution had placed an early bet; its graduates came back to place another.

The neat version of Juvatkar's story contains an acquisition, an accelerator and a large funding round. The more revealing version contains a professor finding travel money, friends learning how to disagree, and engineers permitting their favorite architecture to lose. Nanonets may be in the business of removing manual work. Its co-founder's education was built from doing a great deal of it.