TechCrunch has exclusively learned that Rohit Patel is leaving Meta after an eight-year tenure developing AI. Patel, who previously served as a director at Meta Superintelligence Labs, has unveiled Ndoc, a new startup aiming to be an AI-native alternative to the Microsoft Office suite or Google Workspace.
Patel’s departure follows a series of other prominent exits from Meta, including Chief Global Affairs Officer Joel Kaplan, AI researcher Andrew Tulloch, and Emily Dalton Smith, Head of Product for AI-for-Work Emily Dalton Smith.
Ndoc was conceived by Patel as more than a simple AI add-on to current workflows. Its ambition is to consolidate the disparate methods of work—spanning applications such as Excel, Word, PowerPoint, Tableau, Replit, and the common “ChatGPT copy-paste” approach—into a single, integrated workspace accessible to all collaborators, even those not directly subscribed to Ndoc.
Ndoc unifies three core components within a single interface. The Stage tab serves as the hub for presenting key outcomes, such as slide decks, dashboards, simulations, or mini-applications. The Narrative tab provides analyses, complete with numbered citations, automatically generated figures, and explains the rationale behind the data. The Model tab contains the foundational datasets and Python scripts.
Ndoc’s architecture is a testament to Patel’s professional history. At Meta, he specialized in reinforcement learning—an AI methodology where agents learn optimal decision-making through iterative experimentation to maximize rewards—and assessed the reliability of AI outputs. Patel contends that if AI is to generate the spreadsheets and presentations crucial for corporate decision-making, its work must be verifiable. This is why users can inspect the Python scripts that underpin the documents.
Patel stated, “Even without Python knowledge, you can query [the AI chat] to clarify, ‘What is the function of this script? Which line contains my formula?’ This allows you to confirm the formula’s accuracy.” He emphasized, “There are no hidden elements here.”
Functionally, users can input a dataset and direct Ndoc to analyze it and generate a descriptive narrative of the results. Additionally, document-specific parameters like audience, tone, and memory can be specified. Patel noted, “It retains memory across your individual use, the specific document, your team, and your entire company.”
Patel further explained that Ndoc is constructed and optimized using a combination of open-source models. He added that, due to its ground-up development, the platform’s operational costs are significantly lower compared to solutions relying on cutting-edge frontier AI models.
Reflecting on feedback from early trial users, he remarked, “I anticipated excitement over that aspect, but people aren’t concerned with it.” Instead, he found, “What truly excites people is the ability to manage their entire workflow in a single location.”
Patel’s broader vision involves reimagining the very concept of a file. He states that documents can be exported as a standalone “.ndoc.html” file, viewable in any web browser. This format, he intends to open-source, enabling other developers to create compatible applications, thereby liberating users from the vendor lock-in often associated with Microsoft or Google for decades.
His decision to open-source this endeavor was a key factor in his departure from Meta to pursue this venture independently. He noted that within a large corporation like Meta, achieving consensus on an open-source format would inevitably involve “hundreds of individuals and their perspectives on the format’s design.”
Patel’s impetus was speed. He articulated, “Humans are limited to a human pace, whereas… I can operate at the speed dictated by AI capabilities.”
Aiming to establish itself as the definitive platform for enterprise operations represents a significant gamble for a founder leading a three-person team without external funding, particularly as numerous other AI laboratories vie for dominance in the identical market.
Major players like OpenAI, Anthropic, Meta, SpaceXAI, and Google have all introduced enterprise AI solutions designed to integrate with Google Workspace, Microsoft 365, and other business applications, enabling their AI models to function directly within these environments. Furthermore, OpenAI recently debuted Space, a collaborative workspace within ChatGPT, presenting a more direct challenge to these existing tools.
Patel recognizes Space as a potential competitor but argues that these labs are largely misinterpreting the core problem.
He criticized, “Companies are essentially bolting AI onto spreadsheets to auto-populate cells. That approach frankly disheartens me. It’s akin to discovering electricity and then deciding its primary application is to illuminate the oxen tilling your fields. No, you should invent a tractor.”
However, there is a logical reason why AI labs integrate their models into existing user workflows: altering established habits and overhauling entire workflows is challenging. Patel acknowledges this difficulty, which is why he incorporated connectors, allowing customers to seamlessly import current data from their familiar tools.
He stated, “My standard was to never launch a product I didn’t genuinely enjoy using. I have highly optimized AI workflows for all my tasks. If I begin using this new tool and find myself reverting to my old methods, it signifies that I haven’t developed a sufficiently good product… I have a collection of discarded tools that failed to meet that criterion.”
Patel applies a similar benchmark for achieving user engagement before seeking investment. He explained, “I refrained from accepting funding until I had tangible evidence that customers genuinely appreciated the product.”
Currently, Ndoc is undergoing an enterprise pilot program and making an alpha version available to a select group of customers from its waitlist. While registration is free, users will incur charges based on usage and tokens. Patel anticipates a beta release in the upcoming weeks, with a comprehensive launch scheduled for the following year.
Patel prioritizes rapid iteration over launching a perfect product from day one.
He asserted, “Execution is remarkably inexpensive… therefore, proceed with a less-than-optimal plan and refine it subsequently. By the time one devises the ultimate plan, another will have already completed the execution.”