In 2005, artificial intelligence was a respectable academic interest, though not yet a universal business plan. At Stanford, Vijay Krishnan noticed that the fashionable computer-science specialties were databases and systems. He chose machine learning. It offered the mathematical depth he liked, with its probabilities and statistics, but it could also do something in the world: classify an email, improve a search result, make a useful guess.
The choice carried a quiet piece of family advice. As an undergraduate at IIT Bombay, Krishnan had been attracted to graph theory and approximation algorithms, subjects elegant enough to live almost entirely in the mind. His father nudged him toward applied work. Machine learning became the compromise, if compromise is the right word for finding the field that would occupy the next two decades of your life.
Today Krishnan is the co-founder and chief technology officer of Turing. His compressed biography for himself is “machine learning researcher turned 2x AI entrepreneur.” It is accurate, brisk and a little unfair to the distance contained inside it. The machines have changed beyond recognition. Krishnan’s central concern has not. He has kept working on ways to extract meaning from human behavior, test it, and turn it into decisions that software can make.
One question, four applications
The useful side of the equation
Krishnan began machine-learning research at IIT Bombay in 2003, working with professors Soumen Chakrabarti and Sunita Sarawagi. His papers ranged across text categorization, question answering, entity extraction, web-spam detection and privacy-preserving data mining. At Stanford, where he completed a master’s degree in computer science in 2007, the work continued. So did an important relationship: he met Jonathan Siddharth, who would become his co-founder not once but twice.
After Stanford came Yahoo, where Krishnan worked in machine learning and data mining. It was a short stay and a useful bridge. Search engines had already made machine learning part of ordinary life, though few users called it that. Results were ranked. Advertisements were selected. The grand promise of intelligence arrived disguised as a better ordered page.
In 2008, Krishnan and Siddharth started the company that became Rover. Its first idea was personalized search: helping people recover pages they had seen before, back when browser history was more attic than archive. The company then moved into content discovery, recommending something new from what a reader had previously liked. Its product, Flipora, eventually reached 40 million registered users. The premise was intimate but computational: attention leaves a trail, and the trail can predict what may deserve attention next.
Rover was acquired by the advertising technology company Revcontent in 2017. Krishnan became its senior vice president of data science, then spent several months as an entrepreneur in residence at Foundation Capital. He and Siddharth had finished one long project together. Their next company would begin with a lesson from building it.
“Develop some contrarian and correct insights on important aspects of the world… Ride those insights to their full potential.”Vijay Krishnan's stated success mantra
Opportunity had a geography problem
Rover had relied on talented remote engineers. The founders saw the arrangement not as a concession but as an advantage. Skill was widely distributed; access to the best jobs was not. Companies in places such as Silicon Valley complained about a shortage of engineers while capable developers elsewhere struggled to enter the room. This was a matching problem, and matching problems were familiar territory.
Turing launched in 2018 with the goal of making remote hiring less uncertain. Developers completed technical and soft-skill assessments. The platform used the resulting signals to recommend candidates, predict fit and reduce the hours a company spent interviewing. Krishnan described a system that moved from simpler logistic regression models toward nonlinear and transformer-based ones. At one point, Turing maintained a one-to-one ratio between product engineers and members of its wider data organization. Data science was not a department that received questions. It was the grammar in which the company asked them.
Then the pandemic arrived and made remote work ordinary with indecent speed. A market Turing expected to grow over years moved in months. Yet timing alone does not explain the result. The founders had already built around the unglamorous parts: vetting, ranking, prediction, supply and the logistics of collaboration. Everyone discovered video calls. Turing had spent its early years worrying about whether the person on the call could do the work.
The company became a unicorn in 2021. By then, the original thesis looked less contrarian than obvious, which is generally what happens to a contrarian thesis that survives. But a company built around the measurement of technical ability was about to find another use for that machinery.
The experts behind the model
The generative-AI boom altered the demand. Frontier models could absorb vast quantities of text and code, yet progress increasingly depended on something more deliberate: carefully constructed tasks, expert answers, human feedback and evaluations capable of revealing where a model failed. Turing’s network included programmers and specialists whose knowledge could be organized for that work. The talent cloud could become part of the training loop.
Krishnan draws a firm distinction between data that is plentiful and data that is valuable. Synthetic material can be produced cheaply and at extraordinary volume. Human data is slower, scarcer and more expensive. It may also carry much more information per example, particularly when the human is a domain expert and the problem has a verifiable answer. His metaphor is agreeably mechanical: a model company without the human-in-the-loop component is entering a racetrack with three tyres.
The data paradox - conceptual, not to scale
This work also closes a loop in Krishnan’s career. Rover modeled a reader’s interests. The first Turing modeled a developer’s capabilities. The current Turing helps encode what experts know and determine whether a machine has learned it. Each business asks for a careful description of human judgment. Each then asks where that description can make a better decision.
In 2025, Turing announced a $111 million Series E, bringing its stated funding total to $225 million and its valuation to $2.2 billion. The money was framed around AI data, research and real-world applications. A company once easy to describe as a remote hiring marketplace had acquired a more complicated identity: part talent network, part research partner, part applied-AI builder.
A preference for useful intelligence
Krishnan’s public advice tends to resist enchantment. Newcomers to data science, he says, should learn gradient descent, regression and measurement, then understand how those tools connect to a business. Technical leaders can waste impressive amounts of money when they skip either layer. A model may be sophisticated and still solve the wrong problem. A useful problem may be chosen and then measured carelessly. The expensive comedy of enterprise software is that both errors can look excellent in a presentation.
His framework for deploying AI is similarly practical. Look for the intersection of high business value, tasks models can perform reliably now, and situations that do not require perfect accuracy. The last condition matters. A recommendation may tolerate a mistake that a bank transfer cannot. Intelligence in a laboratory earns applause for possibility. Intelligence in a company earns trust by knowing the cost of being wrong.
Outside work, Krishnan lists chess and nonfiction reading among his interests, especially books about startups, economics, finance and rationality. Chess invites an easy metaphor, so it is best approached with suspicion. Still, there is something positional in his record. He does not appear to have leapt from trend to trend. He has occupied the same square and watched its importance grow: the place where mathematics meets an untidy human decision.
The machines kept changing. The assignment did not: find the signal in human judgment, then make it useful.
IIT Bombay gave Krishnan its Young Alumni Achiever Award two decades after his first research there. By then, machine learning had traveled from specialist elective to corporate imperative. Turing, meanwhile, was helping build datasets and evaluations for models that could write code, reason over technical material and use software. The student who wanted mathematics with a practical outlet had acquired a practical problem large enough to contain almost every branch of mathematics.
There is a temptation to tell any long AI career as prophecy. It makes for a handsome myth and terrible history. Krishnan did not need to foresee chatbots, transformers or a pandemic in 2005. He needed only to notice that learning systems were mathematically interesting and practically consequential. Later, he and Siddharth needed to notice that good engineers existed beyond the hiring radius of good companies. Later still, they needed to see that capable models require structured human expertise.
The aspiration now is larger, but the method remains recognizable: choose an important mismatch, measure it carefully, and build the machinery that lets each side find the other. People and content. Talent and opportunity. Expertise and models. For a man interested in contrarian ideas, Krishnan has built his career on a wonderfully conventional demand - that technology should prove useful.
Research figures and company milestones are current through September 2026. The data graphic expresses the qualitative relationship Krishnan describes and is not a measured dataset.