Devvret Rishi entered the conversation with a disadvantage unusual among prospective co-founders: he had been recommended for his doubts. He was working at Google, had watched attempts to make machine learning easier, and had reasons to be wary. Piero Molino and Travis Addair were exploring a company built around precisely that ambition. A mutual contact thought they should meet. In technology, even skepticism can qualify as a useful introduction.
Rishi tried the open-source projects. Molino and Addair explained their approach. His objections gave way to interest. Addair remembered another detail: Rishi arrived with a presentation about how the technology might become a business. The conversation had acquired a different kind of participant. Here was someone prepared to question the premise and then do some homework about selling it.
Predibase began in 2021. Rishi became its chief product officer and later its CEO. By 2026, following the company’s acquisition by Rubrik, he was general manager of AI at the security and data company. The route between those titles passes through a problem familiar to anyone who has watched an impressive demonstration and then been asked to use the software on Monday.
A model can perform beautifully in a demonstration. A working product must survive the data, the budget and the person whose job depends on its answer. Rishi’s public work keeps returning to the distance between those two experiences.
The long way from a model to a working product
His formal preparation was in computer science. He earned bachelor’s and master’s degrees at Harvard, specializing in machine learning. There was a small venture into representative politics, too: in 2014, he was elected to the university’s Undergraduate Council for Pforzheimer House. Before enterprise customers, a constituency.
He joined Google in 2016, initially working on Firebase, the developer platform Google had acquired. His subsequent work included Google Research, Google Assistant and the organization that became Vertex AI. He was also Kaggle’s first product lead. Those assignments put him near several different ways people encounter machine learning: as researchers, developers, cloud customers and members of a community trying things for themselves.
Kaggle made participation visible. People could work with datasets, compare approaches and learn by doing. An enterprise attempting to put a model into a business process faced further obligations. Somebody had to connect the data, manage training and keep the deployed system running. Rishi had seen both the appetite for machine learning and the effort required to make it operational.
Predibase brought together complementary experiences. Molino had created Ludwig at Uber, giving users a way to specify machine learning tasks through configuration. Addair brought experience with distributed training and Horovod. Stanford professor Chris Ré was also a co-founder. Rishi contributed the perspective of a product manager who had watched users encounter the machinery.
The early proposition was declarative machine learning: describe the task and let the platform handle more of the implementation. Its attraction was quite specific. Users wanted easier access without surrendering the ability to change what was happening underneath. Convenience has a habit of arriving with a locked cupboard; this approach aimed to keep the cupboard accessible.
When Predibase emerged from stealth in May 2022, it announced $16.25 million in combined seed and Series A funding, led by Greylock. Molino was CEO and Rishi was product chief. Rishi subsequently took the CEO role, while Molino became chief scientific officer. A founding photograph freezes the people; the job titles continue moving.
A small model with a specific job
As language models attracted attention, Predibase’s work increasingly concerned customizing open models and serving them efficiently. For Rishi, that meant questions about the actual task a business needed done. A company extracting fields from documents has a different assignment from someone asking a general chatbot to hold forth on every subject under the sun.
In April 2024, Rishi was among the authors of the LoRA Land technical report. The team evaluated 310 fine-tuned models, combining 10 base models with 31 tasks. The report described gains on those evaluations from adapting models to specific work. Its comparison with GPT-4 belonged to those tasks and that experiment. A benchmark has borders, even when a headline is tempted to apply for a passport.
Task-specific evaluations, rather than a claim of universal superiority.
The research also addressed serving. LoRAX allowed multiple adapted models to share base model weights on a GPU. The LoRA Land application hosted 25 specialized Mistral-7B models on a single NVIDIA A100 with 80GB of memory. The distinction between the 310 evaluated models and the 25 hosted models matters: one describes an experiment’s scope, the other a deployment.
There is an economic argument inside that engineering. If different jobs require different adaptations, operating each as a wholly separate system can become expensive. Sharing the underlying resources gives specialization a more workable cost structure. The question becomes how much a particular job needs, and how efficiently that need can be met.
That emphasis suits someone who came to the founding conversation asking how a project becomes a business. The software must be useful enough to buy, and practical enough to keep using. A clever model still has to pay its computing bill.
An extra digit can spoil the whole party
Rishi has named information extraction as his personal favorite among common production language model use cases. The task sounds almost defiantly ordinary: take a document and pull out the fields somebody needs. Names, locations, prices. Put them into a structure another system can read. Much of office life consists of moving such details from one place to another without damaging them.
His example of what can go wrong is wonderfully economical. A document says $10. A model extracts $110. One extra character changes the commercial meaning. The sentence may be fluent and the output neatly formatted; the price is wrong. This is why he discusses output constraints, fine-tuning and validation together.
“think boring”
Devvret Rishi, inviting readers to guess a common production use case
The invitation is funny because the surrounding industry rarely encourages restraint in its choice of subject. Yet repetitive document work presents a serious opportunity. It has volume, an existing cost and a result people can check. The sparkle of the demonstration matters less than whether the fields arrive correctly for the next stage of the process.
His conversations with practitioners have also happened away from the product screen. In 2024, he described dinners co-hosted with Greylock and Galileo where leaders compared views on production language models. The questions were concrete: could open models deliver the required performance, and would the cost make sense? Dinner supplied a setting in which technical preferences had to meet business requirements.
Another public recollection concerned Greylock investor Saam Motamedi, who asked Rishi what was overhyped and underhyped in machine learning. Rishi remembered finding the question difficult on the spot. It is a useful question to remain slightly uncomfortable answering. The answer can change as the technology meets users.
The next meeting has 180 people in it
On June 25, 2025, Rubrik announced an agreement to acquire Predibase. The announcement connected model customization and serving with Rubrik’s governed enterprise data. Rishi’s stated aim was to lower the barriers between an idea and AI running in production. The deal put that work inside a company already concerned with what happens when important systems go wrong.
His subsequent position as Rubrik’s general manager of AI brought another round of listening. He said he spent four months meeting 180 executives after joining. Their concerns focused his attention on the risk of autonomous systems. Once an agent can take action, the business must decide what access it should have, how its behavior is observed and what happens after an error.
Rishi has emphasized monitoring, controls that operate while agents run, and ways to reverse mistakes. Those priorities describe a changed relationship with software. An answer can be read before someone acts on it. An autonomous workflow can make a change before anyone has had the chance to inspect it. Permission and recovery become part of the product’s everyday behavior.
By July 2026, he was discussing agent access and governance on Eye on A.I. In September, his published writing included an article about AI coding tools and shadow IT. The public subjects had moved from making models accessible to managing what people and agents connect to them. His present work follows the consequences of wider adoption.
The questions keep changing
There is a recognizable continuity here, without pretending the career was planned backward from its current title. At Google, Rishi worked on products that brought technology to developers and practitioners. At Predibase, he helped turn open-source foundations into a commercial system. At Rubrik, he works on the controls enterprises need as AI gains the ability to act.
The founding anecdote remains the most revealing one. He had reasons to doubt the proposition. He investigated it. He changed his mind and arrived with ideas about the business. That sequence leaves room for both judgment and curiosity, qualities that become particularly useful in an industry capable of replacing its central question between meetings.
For the person waiting on an extracted price, a deployed model or a safely completed workflow, the measure is less theatrical. Did the system do the work? Can it do it again? Rishi has followed those questions through several versions of AI. There is plenty left to ask.