In October 2020, Jay Nanduri announced that he had left Microsoft. He had spent 22 years there. His farewell was affectionate: the company had shaped him as a developer, a leader and a friend. Then came the practical detail. He was helping start Truveta, he would be its chief technical officer, and he had positions to fill. A considerable stretch of professional life ended with gratitude. The next one began with recruiting.
There is something revealing about that sequence. The announcement gave his colleagues their due, then invited people to build something with him. It offered no elaborate reinvention of his personality. He was taking an engineer’s experience into a new company, where a title would have to be followed by a team and a working product.
By the autumn of 2026, his assignment had expanded again. Nanduri and fellow co-founder Ryan Ahern had become Truveta’s interim co-CEOs. Between those two announcements lies a founder’s story. But the longer route to them passes through a virtual street, a checkout counter and a recurring question about software: what will happen when people actually use it?
Before the agents, a virtual street
Go back to February 2006. Microsoft’s Virtual Earth team was putting an interactive street-side preview online. Users could explore imagery from Seattle and San Francisco, turn their view around and look at a city from something closer to pavement level. The feature went from concept to launch in a few weeks.
Nanduri was among the engineers named for the late-night development sessions that helped get it out. Sean Rowe, the project’s program manager, credited him alongside Brad Snow, Chandu Thota, Eyal Ofek and Rick Welsh, with support from the wider team and its operations crew. It was a collective effort, recorded with the slightly breathless pleasure of people who had just shipped.
The preview also asked users some admirably basic questions. Could they stay oriented? Was navigation easy? Which cities should come next? Behind the clever imagery was the possibility that someone might become thoroughly lost while sitting perfectly still. The team needed people outside the development room to try it.
That early episode makes a useful entrance into a career that later included leadership in Bing Mobile and Bing Maps. Search and mapping put an engineer close to a person trying to get somewhere. An ingenious system still had to help that person take the next step.
The trouble with saying no
Commerce introduced a different navigation problem. A customer could reach the checkout, submit a payment and find the way forward blocked. Sometimes the system had caught a fraudster. Sometimes it had caught a customer.
At Microsoft, Nanduri helped develop fraud protection as the company’s own online commerce grew. The earlier approach relied on rules, including whether the geography of a transaction appeared plausible. The team then examined the customer journey and used machine learning to improve decisions. Even an approved transaction could encounter a bank that lacked the same information and declined it.
This is the awkward charm of real systems: they contain other systems. A merchant’s decision was only one part of a purchase. The work had to account for customers, banks and the people handling manual reviews. Improving one checkpoint could leave the overall experience stubbornly unimproved.
By April 2019, Nanduri was introducing the public preview of Dynamics 365 Fraud Protection. Microsoft had built the technology to meet its own needs and was making it available to other merchants. Customers could examine historical transactions before committing to a more involved evaluation. They could then compare live assessments against their existing system.
That sequence gave buyers a way to investigate the product before depending on it. It also reflected the central balancing act: reduce fraudulent purchases while allowing legitimate ones through. A fraud system can make its own life wonderfully simple by declining everybody. The shopkeeper is unlikely to applaud.
The project brought Nanduri together with Anand Oka, Yuting Jia, John Beaver and Yung-Wen Liu on Microsoft’s 2019 Franz Edelman finalist team. The recognition concerned an application of analytics and operations research, and Nanduri was inducted as a laureate. Oka would appear again in a later chapter of his working life.
In a paper published in January 2020, the five collaborators described models for historical and emerging fraud patterns, alongside decisions that accounted for what other parties would do next. Their reported results covered the 2016-2018 implementation period. This was engineering with consequences measured in completed transactions and reduced losses.
Incremental fraud costs saved over two years through reduced fraud losses and manual reviews.
A result reported for Microsoft’s system and team, rather than an individual attribution.The bill arrives after the demo
By the time he left Microsoft, Nanduri had become a Technical Fellow. His experience included commerce platforms, financial services and supply chain work. The job titles changed across that period, but operating software at scale repeatedly put him in contact with constraints that an appealing demonstration can politely leave offstage.
His later writing at Truveta spends time on those constraints. In July 2024, he discussed the development and operating costs of large language models. He described using networks of smaller, specially tuned models for complex tasks, and choosing modeling approaches according to the work they needed to perform. Hardware availability and inference costs belonged in the calculation.
There is a certain relief in encountering a technology executive willing to talk about the bill. Every additional capability has to run somewhere. Every repeated task consumes time and resources. Nanduri’s account treats those expenses as design considerations, alongside the quality of the answer.
Quality has its own complications. A model adapted for a new task may share components with an older one. In his June 2024 discussion of accuracy, he explained why the earlier task needs to be checked again when those components change. A new capability can disturb something that already worked. Monitoring also continues after deployment, because incoming data can change.
“An AI model is only as good as the data it’s trained on.”
Jay Nanduri, Seattle University panel, 2024
His December essay extended the argument to reusable networks of specialized AI agents. He described orchestration, transfer learning and controls over which versions an application could use. The ambition was to let useful knowledge serve several applications while preserving control over how it was deployed.
His writing also puts people inside the process. In August 2024, he described human experts supplying reference judgments, monitoring errors and retaining the ability to override AI actions. Low-confidence outputs could be withheld. An automated system still needed a route for uncertainty, and someone responsible for handling it.
By October, he was addressing adoption: documentation, worked examples, videos and customer boot camps. These are familiar instruments of patient explanation. A new tool asks its users to learn something, too. Giving them a place to begin is part of delivering the tool, even if it makes a less dramatic announcement.
Read together, these ideas give his AI work a recognizable engineering texture. Models need coordination. Changes need checks. Existing users still need their work to succeed when the next feature arrives. The exciting part of a new system occupies the same calendar as its upkeep.
A book with somewhere else to go
In 2026, Nanduri and Oka published Agentic AI Unleashed. The collaboration connected two people who had already worked together on Microsoft’s fraud systems. The book carries Nanduri’s longer name, Jayaram, on its cover and takes readers through designing, building and deploying autonomous AI systems.
Its contents move through tools, memory, development, testing, observability and deployment. Those subjects give the project a practical center of gravity. Someone has to understand how the pieces fit together, then keep track of what they do once they are running.
The authors launched the book at Truveta’s Annual Homecoming. Copies went to employees, and the pair signed as many as they could. Nanduri said they hoped the book would reach classrooms as well. He also announced that all author proceeds would support nonprofits advancing education and access to opportunity.
Education already had a place in his life. Givology’s 2008-2009 annual report lists him on its board of directors, identifying him then as a Microsoft principal development manager. Years before the book, he was connected to an organization supporting education through a volunteer network.
That gives the pledge a history. The book gathers professional knowledge into something another person can study. Its proceeds have a stated destination beyond the authors. For a career so often concerned with making shared resources useful, it is a fittingly concrete arrangement.
A bigger job, with another name beside his
The September 2026 leadership change brought Nanduri into Truveta’s top job alongside Ahern as Terry Myerson retired as CEO. Nanduri continues to lead engineering, product and the data systems behind them. Ahern handles the commercial and delivery side of the business. The company had also opened a development center in Hyderabad that July, extending its engineering organization.
“Ryan takes care of demand,” Nanduri explained. It is a compact description from someone accustomed to thinking about how parts of a system relate. Sharing the top job makes the connection between their responsibilities particularly visible.
Twenty-two years at Microsoft, followed by founding a company, writing a book and taking on an expanded leadership role, could be arranged into a tidy procession of titles. The more interesting record is full of working relationships: engineers finishing a mapping preview, colleagues balancing commerce decisions, a returning co-author, another founder sharing responsibility.
The next chapter gives Nanduri more of those connections to manage. His career has already supplied plenty of practice with the moment when the demonstration ends, someone tries the product, and the work becomes specific.