Meta AI director leaves to launch Ndoc workspace

From Meta to a new startup

Rohit Patel is leaving Meta after eight years of developing artificial intelligence for the company. During his time there, he served as a director at Meta Superintelligence Labs, where he focused on reinforcement learning. This is a type of machine learning where AI agents learn to make decisions by trying different actions and maximizing rewards.

Patel is joining a list of high-profile departures from the company in recent months. Other notable exits include Chief Global Affairs Officer Joel Kaplan and AI researcher Andrew Tulloch. His move marks a transition from corporate research to building a startup that challenges the current status quo of enterprise software.

He is launching Ndoc, which he describes as an AI-native replacement for the Microsoft Office suite and Google Workspace. Rather than simply adding a chatbot to an existing spreadsheet or document, Patel wants to build a workspace where AI is the core foundation. He aims to solve the problem of people having to copy and paste information between many different disconnected tools.

How Ndoc unifies the workflow

Ndoc is designed to bring three distinct components into one layout to streamline how teams handle information. The Stage tab is intended for presentations, housing things like slide decks, dashboards, and simulations. The Narrative tab is where the 'why' behind the numbers lives, featuring generated figures and analyses with numbered citations.

The third component is the Model tab, which stores the underlying datasets and Python scripts used to generate the work. This design allows users to upload a dataset and ask the AI to perform an analysis and write a narrative about the findings. Users can also provide specific context, such as the intended audience, the desired tone, and the company's memory.

Patel emphasizes that the platform is built to be transparent rather than a 'black box.' Because the underlying Python scripts are visible, users can ask the AI to explain specific lines of code or formulas. He believes that if AI is going to build the documents used for corporate decision-making, people must be able to verify that the work is accurate.

Technical choices and costs

The startup is taking a different technical approach than many of its larger competitors. Ndoc is built and fine-tuned using a mixture of open-source models rather than relying solely on expensive frontier models. Patel notes that this choice makes it considerably less expensive to run the platform for customers.

While he initially thought customers would be excited about the lower costs, he found that users are more interested in the end-to-end experience. They want to see all their work in one place rather than worrying about the underlying infrastructure. To help with this transition, he included connectors so customers can import data from the tools they already use.

Patel is also focused on creating a new file format called .ndoc.html. He plans to open-source this format so that other developers can build apps that work with it. His goal is to prevent the kind of ecosystem lock-in that has occurred with Microsoft and Google over the last several decades.

The competition and strategy

Ndoc faces significant competition from established giants like OpenAI, Anthropic, and Google. These companies are currently releasing enterprise tools that allow their AI models to work directly inside existing apps like Microsoft 365. OpenAI recently launched Space, which is a shared workspace inside ChatGPT that competes directly with the tools Patel is trying to replace.

Patel argues that these larger labs have the wrong idea about how AI should be used in the workplace. He compares attaching AI to old spreadsheets to putting lights on a bull to till a field, suggesting that the underlying tool should be replaced by a tractor. He believes the current trend of just filling cells in a spreadsheet kills the soul of the work.

Despite the competition, Patel is moving quickly to gain traction before seeking funding. He is currently conducting an enterprise pilot and opening an alpha version to a limited waitlist. He intends to release a beta version in the coming weeks, with a full launch scheduled for next year. He believes that executing on a suboptimal plan now is better than waiting for a perfect plan while others move ahead.

A high-stakes bet

Starting a company with a team of only three people and no funding is a significant risk in the current AI landscape. Patel is betting that a specialized, AI-native workspace can win over users who are tired of fragmented workflows. He has already discarded several tools that did not meet his personal bar for optimization and usability.

His philosophy is rooted in the idea that execution is the most important factor for success. He wants to prove that people love the product before he takes any outside money. By focusing on a unified experience and an open format, he hopes to create a new standard for how enterprise work is conducted.

If Ndoc succeeds, it could change how companies organize their internal data and communication. Instead of moving data between different apps, the data would live in a single environment where the AI understands the context of the entire project. This shift would move the needle from using AI as a helper to using AI as the primary architect of the workspace.

The path forward

The next phase for Ndoc involves moving from the current alpha stage to a broader beta release. During this time, Patel will be looking for more proof of product-market fit among enterprise customers. The goal is to demonstrate that the unified Stage, Narrative, and Model tabs provide a superior experience to the current status quo.

Observers will be watching to see if the .ndoc.html format gains any traction with the developer community. If the format becomes a standard, it could provide a way for different apps to communicate without being tied to a single provider. This would be a major shift in how software interoperability is handled in the AI era.

For now, the startup is focused on the immediate goal of making the tool indispensable for daily tasks. Patel wants to ensure that users don't just try the tool once but keep coming back to it. His success depends on whether his vision of a 'tractor' can replace the 'bulls' that currently dominate the corporate office.

Verification and transparency

One of the core pillars of Ndoc is the ability for non-technical users to audit AI-generated content. By housing the Python scripts in the Model tab, Patel allows anyone to ask the AI to explain the logic behind a specific calculation. This transparency is intended to build trust with corporate users who are often wary of AI hallucinations.

The platform also maintains a persistent memory across the user, the specific document, the team, and the entire company. This means the AI doesn't have to be