When Devavrat Shah arrived at Stanford in 1999, the startup question was already waiting for him. Someone outside the university heard that he was a computer science PhD student and immediately wanted to know the name of his company. In that particular time and place, a doctorate apparently required a business plan.
Shah tried the business route. During graduate school, he took leave to work at a networking startup and developed a memory-circuit solution. Then the interesting problem gave way to the work of execution. “I realized that the innovative phase is over,” he recalled. He returned to Stanford.
The decision has acquired a pleasing irony. Shah would go on to co-found Celect, bought by Nike in 2019, and Ikigai, acquired by Celonis in 2026. He kept his academic career, too. Looking across those choices, the recurring attraction seems to be a problem with enough uncertainty left in it to reward another idea.
The route to class went past the coffee stall
Before Stanford came IIT Bombay, where Shah completed his computer science and engineering degree in 1999 and received the President of India Gold Medal. His recollections of those years are considerably less ceremonial: hostel conversations late at night, chai at Chinkos, and nights in the computer science labs, whose air conditioning merited a fond mention.
He also remembered taking the “shortest” path between lectures through the coffee stall. Any campus map that leaves out the refreshments is clearly missing a variable. There were modaks at his host professor Siva Kumar’s home, too. The memories give an otherwise formidable academic record the ordinary texture of student life.
Shah grew up in Vadodara, in a family with both academic and business traditions. His father taught mechanical engineering; his grandfather taught chemistry. His later career would leave room for research and company-building in the same working life.
The network learns to take turns
He earned his Stanford doctorate in 2004, did postdoctoral work in statistics, and joined MIT in 2005. His dissertation concerned switch algorithms: part of the machinery that helps information move through networks. It received first place in the 2005 George B. Dantzig Dissertation Award.
The vocabulary can sound remote from daily life. The difficulty is familiar: many participants need access to something scarce. A communication network has to coordinate their demands without wasting the resource in collisions or delays. Giving every participant a central referee can introduce costs of its own.
Shah’s research has examined how local exchanges can produce useful behavior across a network. His work on gossip algorithms studies the spread and combination of information through those exchanges. His work on belief propagation connects message-passing methods with optimization. In 2025, ACM SIGMETRICS recognized his contributions with its Achievement Award.
Earlier recognition included an NSF CAREER Award in 2006, the first ACM SIGMETRICS Rising Star Award in 2008, and the INFORMS Erlang Prize in 2010. The honors came from communities concerned with computing, networks and applied probability.
That history helps explain his interest in methods that work under tight computational constraints. The elegance of an answer matters; so does whether a real system can afford to calculate it.
A sale leaves out the alternatives
Human choice gives the mathematician another version of the incomplete-information problem. A shop can record what a customer bought. The receipt says much less about the alternatives on the shelf or the item the customer wanted but could not find.
Shah’s work on social data processing pays attention to comparisons. People often find it easier to express a preference between two options than to attach an absolute score to one. The surrounding assortment can change what a purchase tells you. Selling a shirt beside one set of competing products carries different information from selling it beside another.
His wider argument is that engineers need to account for how people actually choose when designing systems people will use. In that view, behavior belongs inside the model from the beginning. Otherwise, the system can be mathematically tidy while making unreasonable demands on everyone expected to operate it.
A business forecast is full of these human traces. Behind its numbers are offers, decisions and circumstances. Learning from the table requires some respect for how the table came into existence.
Nike buys the forecast
Celect turned research on choice into a retail business. Shah co-founded the company with MIT colleague Vivek Farias. The practical questions involved inventory: which products belonged where, and what might customers want when they arrived?
On August 6, 2019, Nike announced that it had acquired Celect. The attraction was demand sensing and predictive analytics, including local forecasts that could help allocate inventory across shopping channels. A product needed to reach an appropriate place before the customer’s interest became a missed sale.
The acquisition brought Celect’s team into Nike’s global operations organization. The co-founders would remain tenured MIT professors and consult for Nike. Research and its commercial application continued alongside each other, with different responsibilities attached.
It is a useful episode in Shah’s story because the object of the mathematics becomes tangible. Inventory takes up space. Customers make choices. Decisions about quantities and locations have consequences well before anyone opens a spreadsheet to review the quarter.
A classroom with the door open
The teaching thread continued. Shah’s MIT graduate course, Algorithms for Inference, introduces probabilistic models represented through graphs. Its materials are available through OpenCourseWare, including lecture notes, assignments and exams.
The course connects ideas used across machine learning, communication, computer vision and control. It offers another way of seeing the continuity in his work: a common set of methods can travel between problems that look unrelated when described by their industry labels.
He also serves as faculty director of the MicroMasters program in Statistics and Data Science at MIT’s Institute for Data, Systems, and Society. Teaching gives the career another audience beyond research collaborators and enterprise customers. Someone opening the course materials can follow the reasoning at its own pace, before a product demonstration or a forecast ever enters the conversation.
A professor, a former student, a different model
In 2019, Shah co-founded Ikigai with Vinayak Ramesh, his former MIT student. Their company focused on structured business data, including tables and time series. Much of a company’s operating history lives in that form: rows, columns, dates and inconvenient blanks.
Ikigai built around large graphical models, or LGMs. In Shah’s explanation, they learn relationships within tabular information, including relationships across separate tables. The intended applications include forecasting and exploring different decisions. A conversational interface may help a person ask a question; the numerical relationships still need a suitable model.
Asked how a customer could check generated values, Shah offered a practical answer: remove some known data, ask the system to reconstruct it, then compare its answers with the values held back. That test, cross-validation, makes room for skepticism inside the demonstration.
The important distinction is between a capability and evidence that it works on a particular problem. Filling a blank confidently is easy to display. Measuring whether the reconstruction deserves confidence is a more demanding conversation, especially when people will act on the result.
“Persistence, hard work and learning from everyone”
Devavrat Shah / on his success mantra
The forecast needs people in the room
Shah has described planning as a team sport. In a public post, he argued for bringing participants together around data so that information does not disappear into separate spreadsheets, messages or conversations. Anyone who has attended a meeting devoted to discovering who knew what will recognize the ambition.
His discussion of consumption-based pricing makes the point more concrete. Forecasting total activity can be easier than predicting which customer, channel or sales representative will account for it. Looking across related groups can reveal patterns that remain elusive when each account is considered alone.
He presents AI as a way to help those groups learn from one another and build more trust in the forecast. It is a practical social objective attached to a technical method. The calculation has to survive the meeting in which people decide whether to use it.
The next decision, at Celonis
Celonis announced its agreement to acquire Ikigai on May 12, 2026. Shah would become Chief Scientist, Enterprise AI, and Ramesh would become Field CTO. By July, the acquisition and Shah’s new role were confirmed alongside his continuing MIT work.
The combination brings Ikigai’s forecasting and simulation work to Celonis’s models of business operations. The aim is to connect an understanding of how work happens with a way to explore what might happen next. That gives Shah’s research a larger setting in which to meet the complications of actual organizations.
In July, he described the attraction of concentrating on structured and time-dependent data: “A narrower focus comes with sharper technology.” It is an apt description of the career as well. The applications have widened while the questions have remained specific.
He has also argued for introducing enterprise AI around functional decision-making roles. That keeps attention on a recognizable responsibility within an organization, rather than an arbitrary collection of automated actions. Someone still needs to understand what decision is being made and how its consequences connect to the rest of the business.
From the early networking problem to the next planning model, Shah keeps returning to information that is scattered, limited or incomplete. His working life has made room for both the pleasure of finding a solution and the awkward business of putting it to use. The spreadsheet, for its part, continues to offer plenty of unanswered questions.
Follow the questions
Research, conversations and further reading.