On a Stanford web server, in a page so plain it could have been preserved in amber, Arvind Sundararajan describes his career with comic economy. He earned a computer-science master's degree there in 1998, he writes. Before that came a B.Tech from IIT Bombay. Since school, he has worked at “a few different companies in the valley, and started a couple.” The line is technically true in the way that calling the Pacific a quantity of water is technically true.
Those few companies include Google during Gmail's launch years, Uber's autonomous-vehicle operation and three startups. The couple he started include Post Intelligence, acquired by Uber in 2017, and Abacus.AI, where he is CTO and co-founder. In between, there were large advertising systems, social feeds moderated by deep learning, software riding inside experimental vehicles, patents and a long-context language-model paper called Giraffe.
His work has moved with the fashions of computing, but the underlying problem has stayed remarkably still. Complicated systems are useful only when they can be made reliable, scalable and simple enough for somebody else to use. Sundararajan keeps returning to the machinery beneath that simplicity. The interface gets the applause. The infrastructure gets him.
The inbox as an infrastructure lesson
In 2004, Sundararajan joined Google. Gmail was arriving with an offer that sounded faintly absurd at the time: a gigabyte of free storage, search instead of filing, and conversations instead of scattered messages. None of that charm mattered if the backend faltered. Sundararajan became a technical lead for Gmail's backend and was part of its early launch team, helping the service grow through its first tens of millions of users.
He later led the machine-learning and serving system behind AdSense, one of Google's essential advertising businesses. He received a Google Founders Award for each body of work. The pairing is revealing. Email and advertising look like different products, but from an engineer's chair both demand relentless serving, rapid decisions and a low tolerance for drama. Users may forgive a clever system. They do not forgive a missing message or the wrong result delivered slowly.
That period formed the grammar of Sundararajan's later career: data in motion, models in production, and systems expected to work for populations much larger than the team operating them. The central skill was not invention in isolation. It was making invention survive contact with scale.
“Still madly in love with the Internet.”Sundararajan's public profile bio
A founder partnership gets a second act
In 2010, Sundararajan and Bindu Reddy co-founded the company that became Post Intelligence. Its premise was early for its moment: learn how people and brands behave on social networks, then use machine learning to help them decide what to publish. The company also developed Candid, an anonymous newsfeed whose moderation system used deep learning. Years before today's argument over automated moderation became a dinner-table subject, the team was already wrestling with machines that had to interpret messy human expression.
The pair did not merely share a cap table. Reddy worked at the product and business edge; Sundararajan served as CTO and built the technical core. Post Intelligence reached the end point every startup deck quietly desires when Uber acquired it in 2017. The acquisition sent Sundararajan into a quite different environment: the company's Advanced Technologies Group, where software had wheels, sensors and physical consequences.
He led the Autonomy Systems team across onboard software, machine learning and infrastructure. It was a broad assignment. An autonomous-vehicle stack has to perceive, decide and act, while a thicket of supporting systems records, trains, tests and deploys. The old concern with dependable complexity remained, but the computer was no longer safely tucked into a data center. It was moving through the world.
By 2019, Sundararajan and Reddy were ready to build together again. They joined with Siddartha Naidu, known for his work on Google's BigQuery, and founded RealityEngines.AI. The name was grand and a little chilly. In 2020, the company became Abacus.AI, borrowing the oldest familiar calculating tool for a business devoted to modern computation. A machine simple enough for a child to understand made a useful emblem for software designed to conceal a great deal of difficulty.
Fifteen minutes and a difficult promise
Abacus.AI began with an unfashionable observation about fashionable technology. Machine learning might be powerful, but production machine learning was a nuisance. A working model was only one part of the job. Companies also needed data pipelines, cleaning, retraining, monitoring, feature storage and capacity that could expand when predictions arrived in volume. Scarce specialists were being asked to assemble the whole contraption.
The founders proposed to package that work. A customer could bring data and a use case; the platform would help select, train, deploy and operate the model. Fifteen minutes into their first meeting with Index Ventures, the investors were sold. This is either a tribute to clarity or to the excellent timekeeping of venture capitalists. Probably both.
The ambition attracted capital quickly. The company announced a $13 million Series A in 2020, then a $22 million round later that year. In October 2021 it raised a $50 million Series C led by Tiger Global, with Coatue, Index Ventures and Alkeon Capital participating. Disclosed funding passed $90 million. At that point, Abacus.AI said 10,000 users had created more than 30,000 models on the platform.
The hidden-complexity index · conceptual, not financial
The chart is playful, but the progression is real. Each chapter placed more kinds of machinery behind the screen. Email concealed global storage and serving. Advertising added rapid prediction. Autonomous vehicles joined software to uncertain physical environments. General-purpose agents combine models, applications, browsing, code execution and a computer, then invite a non-specialist to ask for an outcome in ordinary language.
The research habit behind the product
Abacus.AI did not abandon research when generative AI rearranged the market. Its work spans forecasting, recommendation, automated machine learning, model fairness and neural architecture search. Sundararajan appears among the authors of Giraffe, a 2023 study of how language models might operate beyond the context lengths used during training.
The team tested methods on LLaMA-family models, built new evaluation tasks and released three long-context models at 4,000, 16,000 and 32,000 tokens. The paper's finding was practical rather than theatrical: among the approaches tested, linear scaling performed best, while a truncated positional basis showed promise. It was work in the old Sundararajan mode, taking a limitation that users feel and treating it as a systems problem that can be measured.
His earlier patents tell a similar story. They concern search over structured data and the addition of attributes and labels to structured data. The nouns have changed since the 2000s. The verbs have not: structure, search, serve, scale.
“Our goal is to make a product which lets non-technical users automate all kinds of knowledge work.”Arvind Sundararajan
Calm code in a noisy industry
There is not a vast archive of Sundararajan holding forth about leadership. What exists is more useful. Rob McCool, who worked with him at Alpiri, remembered him as one of the tiny startup's first hires. In a company that never grew beyond six people, the third person mattered enormously. McCool described Sundararajan as easygoing and enthusiastic, with clean, reliable code and strong design instincts. When requirements shifted and deadlines materialized, he took the chaos in stride.
That recollection does more than furnish a personality adjective. Infrastructure work is an emotional discipline. The system may be frantic; the engineer cannot be. Sundararajan's career repeatedly put him where changing requirements met unforgiving operations, from a young Gmail to an early startup to vehicles attempting to understand a street.
His own public style is notably spare. He posts infrequently. He does not manufacture profundities between product announcements. On LinkedIn, he celebrated Abacus.AI coming out of stealth, its rebrand and its funding, usually by pointing attention toward the team or the work. The understatement of his Stanford homepage feels less like false modesty than preference. There are systems to build, and the internet remains, in his own phrase, an object of affection.
A computer behind the conversation
Abacus.AI's current direction brings the career full circle. Deep Agent is presented as a cloud service with a consumer-grade interface, access to leading language models, infrastructure tools and a complete computer. It is meant to create data-driven sites and applications, make videos, conduct research and solve coding problems. The product promise is no longer merely to build a model. It is to complete work.
For a non-technical user, that shift can feel magical. For an infrastructure engineer, magic is a backlog. Every smooth instruction can trigger model selection, tool use, file handling, browsing, execution, checking and recovery. The more natural the conversation becomes, the more engineering must disappear beneath it.
This is why Sundararajan's path is more coherent than its collection of famous names suggests. Gmail, AdSense, Post Intelligence, Uber ATG and Abacus.AI are not trophies on a shelf. They are variations on a stubborn pursuit: take a system with too many moving pieces and turn it into something dependable enough to be used without ceremony.
The Stanford homepage ends with an invitation to join a high-velocity team building a challenging generative-AI application. The phrasing is direct. No manifesto, no prophecy, no promise that software will abolish work by Tuesday. Only the old bargain of ambitious engineering: the problem is hard, the team is moving, and useful simplicity has to be earned.