TechCrunch has exclusively learned that after 8 years of engaging in AI research and development at Meta, Rohit Patel is about to leave the company. Patel, who once served as a supervisor at Meta Superintelligence Labs, has now founded a startup called Ndoc and positioned it as a native alternative to Microsoft Office suites or Google Workspace.
Patel is the latest among a group of senior executives and researchers who recently resigned from Meta. Those who left before him include the Chief Global Affairs Officer Joel Kaplan, Researcher AI Andrew Tulloch, as well as the Product Manager in charge of AI office products, Emily Dalton Smith.
When designing Ndoc, Patel did not simply involve overlaying AI tools onto the existing workflow. His goal was to replace the fragmented way people worked, which involved constantly switching between Excel, Word, PowerPoint, Tableau, Replit, and the “ChatGPT copy and paste” process. Instead, he aimed to create a unified workspace that could be accessed by everyone, even if colleagues were not registered with Ndoc.
Ndoc Integrates the three parts into a single interface. The Stage tab is used to display key results, such as slides, dashboards, simulations, or mini-applications; the Narrative tab contains analyses, numbered references, chart generation, and content typically used to explain the reasons behind the numbers; the Model tab stores the underlying datasets, and the Python tab holds the scripts.
This design reflects the background of Patel. During his tenure at Meta, he mainly focused on research in reinforcement learning—a method of machine learning where AI agents make decisions through trial and error to maximize rewards—as well as on evaluating the credibility of AI outputs. Patel believes that if AI is to generate spreadsheets and slides for corporate decision-making, there should be someone who can review its work. For example, users should be able to find the Python scripts used to generate these documents.
Patel said, "Even if you don't know a thing about Python, you can still ask the AI chat function to explain, 'Hey, what is this script for? Which line is my formula?' Then you can verify that formula. There's no hidden content here at all."
In practical use, users can upload datasets, allowing Ndoc to analyze them and produce a narrative regarding the results. Users can also add document-specific context, including the audience, tone, and memory information. Patel states: "It will continuously retain this memory at the individual, specific document, team, and company levels."
Patel also indicates that Ndoc is constructed using a variety of open-source models and fine-tuned accordingly. Since this platform was developed from scratch, he claims that its operating costs are significantly lower than those of alternative solutions built on cutting-edge AI models.
“I originally thought this would be the point that people would be most interested in, but everyone didn’t seem to care about it,” he said, referring to the early trial users. “What really excites people is their preference for putting all end-to-end processes in one place.”
Patel's bigger ambition is to redefine the file itself. He stated that documents can be exported as a single “. ndoc.html” file and can be opened in any browser; he plans to open-source this format so that other developers can build applications based on it, without being restricted by Ndoc as many people have been for years by Microsoft or Google.
Part of the reason he wanted to open-source this project was also because of his decision to leave Meta and go it alone. In a large company like Meta, it is obviously necessary to have "hundreds of people and various opinions on how the format should be designed" in order to reach a consensus on an open-source format.
Patel hopes to advance as quickly as possible. "Humans can only move at the speed of humans, but I... can move forward at the speed that AI is capable of."
To become a platform for corporate work is a huge gamble for a founder with just a team of three people and no financing, especially since other AI laboratories are also competing in the same field.
OpenAI, Anthropic, Meta, SpaceXAI, and Google have all launched AI tools that allow users to connect to Google Workspace, Microsoft 365, and other office applications, enabling the AI model to work directly within these tools. Last month, OpenAI also introduced Space, which is a shared workspace within ChatGPT and poses more direct competition to these established tools.
Patel acknowledges that Space may pose a threat, but he believes that the approaches of most of these laboratories are incorrect.
“These companies are basically just taking AI and putting it into spreadsheets, letting it fill in the cells for them,” he said. “That really bothers me. It’s like you’ve just invented electricity, and then you decide the best use for it is to put lights on the bulls that plow my fields. Don’t do that. You should be making tractors.”
However, there are reasons why the AI laboratory attempts to bring the model to places where people are already working. Changing habits and replacing entire workflows is not easy. Patel indicates that he understands this, so he built connectors that allow customers to import existing data from the tools they are already using.
“My standard is that I won’t release anything that I myself wouldn’t like to use,” he said. “I’ve created a highly optimized AI workflow for myself. If I start using this tool but still keep returning to my original workflow, it means I haven’t created a product good enough yet… I have a whole graveyard of tools that don’t meet this standard.”
Before attempting to raise funds, Patel also has similar criteria for gaining customer recognition. “I don’t want to take money before I have some kind of proof that people really like this product,” he said.
As of Tuesday, Ndoc is conducting corporate pilots and has opened the alpha version to a small number of customers on the waiting list. Registration is free, but customers will need to pay for usage and tokens. He hopes to launch the beta version in the coming weeks and plans to officially release it next year.
For Patel, moving forward as quickly as possible is more important than having a perfect product from the start.
“The cost of implementation is extremely low… so let’s implement an imperfect plan for now, and we can discuss further later,” he said. “By the time you come up with the best plan, someone else might have already completed the implementation work.”
