Eight years inside Meta's AI operation. A director's title at Meta Superintelligence Labs. Rohit Patel just traded all of it for a team of three and no outside funding, and his target is the software sitting on nearly every work laptop on the planet.
The product is called Ndoc. Patel describes it as a ground-up rethink of the office suite, built around AI from day one instead of patched with it, and aimed squarely at both Microsoft's and Google's bundles. TechCrunch was first to report it. Starting that fight from zero is a bold choice. So what has he built, and why does he think the giants have it backward?
A short runway in a crowded field
Patel is one of several people to have left Meta recently. Joel Kaplan, the chief global affairs officer, has departed. So has AI researcher Andrew Tulloch. So has Emily Dalton Smith, who ran product for AI-for-work. Nobody has publicly tied those exits into a single story, and they may not be one.
Patel's own reason is oddly specific. He wants to release an open file format, and he says that at Meta, getting one approved would mean absorbing the opinions of hundreds of people. On his own, he can move as fast as the AI tooling lets him. Founders say this a lot. It's also a fair description of how big companies treat anything meant to be a standard.
The field he's entering isn't empty. Every major lab, from OpenAI and Anthropic to Meta, SpaceXAI and Google, now sells enterprise products that wire its models into the apps people already live in, including Microsoft 365 and Google Workspace. Last month OpenAI went a step further with Space, a shared workspace living inside ChatGPT, which looks like a much more direct rival to the old suites. Patel concedes Space could hurt him. He just doesn't think the labs have the right idea about what AI should do at work.
Three tabs and a paper trail
The core of the product is also where Patel's résumé shows. At Meta he worked on reinforcement learning (the trial-and-error method where software agents learn which decisions earn rewards) and on judging whether AI output can be trusted. Ndoc is basically that second obsession turned into an interface.
The workspace has three tabs. Stage is the presentation layer: slide decks, dashboards, simulations, small apps. Narrative holds the written analysis, complete with numbered citations and generated figures, the explanation behind the numbers. Model is the engine room, where the datasets and the Python scripts live.
That last tab is the interesting one. Most AI tools that produce a spreadsheet or a deck hand you a finished artifact and ask for your faith. Ndoc exposes the code that made it. And if you can't read Python, you can ask the built-in chat to walk you through a script, point to the line holding a particular formula, and confirm it does what it claims. Patel puts it simply: "There's nothing hidden behind the curtain here."
Say you're a finance analyst and a chatbot hands you a number. You can trust it or redo the work by hand. A workspace that shows the derivation, in a form a non-programmer can interrogate, adds a third option: check it in five minutes. Whether you'll actually do that is another matter (most of us skim). Still, having the option is a real design choice, and the bolt-on assistants mostly skip it.
There's more. You can upload a dataset, ask for an analysis, and get a written narrative back. You can attach context to a document, such as the intended audience and tone. And according to Patel, Ndoc keeps memory at four levels: the individual, the specific document, the team, and the company. In theory, a quarterly report written for the board shouldn't sound like a Slack update to the design crew.
Colleagues can open and view shared work without signing up for Ndoc themselves. That matters more than it sounds. Every new collaboration tool dies on the same hill: you can't get the whole team to switch on the same day. Letting outsiders in through the side door at least lowers the toll.
One more detail, because it speaks to the economics. Patel says Ndoc runs on a blend of open-source models that he fine-tuned himself, rather than on one of the expensive frontier systems, and that this makes it far cheaper to operate. He assumed early testers would care. They didn't. What excited them was seeing the whole chain, from raw data to final deliverable, in one place. The thing you're proudest of in the engine room is rarely what the customer notices from the driver's seat.
The case against attaching AI to a spreadsheet
Patel is blunt about the labs' strategy. Letting a model fill in spreadsheet cells strikes him as a waste of a new technology, and he reached for a farming metaphor to say so: having just discovered electricity, you wire up lights on the plow when you should be building a tractor. "Get a tractor."
It's a good line. It's also incomplete, and Patel seems to know it. The labs are bolting onto Excel and Word for a reason: people hate changing how they work. Entire companies run on workflows that nobody under 50 can explain and nobody over 50 will touch. Meeting people inside the tools they already use is the lowest-friction way to sell AI to a business.
His answer is a set of connectors that import data from the tools customers already use, so nobody starts from zero. He's also set himself a personal test for shipping. He runs heavily optimized AI workflows for his own work, and if he catches himself drifting back to them instead of using Ndoc, the product isn't ready. By his own account, he has "a graveyard of tools that did not meet that bar." Refreshing, and a little alarming, since it implies several ideas died before this one.
The file is the real bet
Strip away the interface and Patel's bigger ambition is the file itself. Any Ndoc document can be saved as one self-contained .ndoc.html file, which any web browser can open. He plans to open source that format, so other developers can build tools that read and write it without being tied to Ndoc, the way businesses have spent decades tied to Microsoft and Google formats.
This is the part I'd watch most closely. A workspace app is a product. A file format that others adopt is infrastructure, and infrastructure outlasts products. PDF survived because anyone could produce and read one, not because Adobe's reader was lovable. If an HTML-based format can carry the narrative, the data, and the code together in a single portable object, that's a genuinely different idea of what a work document is.
It's also a long shot. Formats win through adoption, and adoption needs a critical mass of other people's software. A three-person startup can publish a spec on a Tuesday. Getting anyone to build against it is a years-long slog, and the incumbents have every reason to sit back and watch.
What to watch
Two things will tell us fast whether Ndoc is a company or a good demo. First, whether the open format attracts any outside developers once Patel actually releases it, since that's the move that would make the product hard to dislodge. Second, how OpenAI's Space evolves. If the ChatGPT-based workspace absorbs the "one place for everything" appeal, Ndoc's main selling point gets squeezed.
There's also an unanswered funding question. Patel is running with no outside money and a tiny team, and nobody has said how long that holds or what customer numbers he wants to hit before raising. Ambition is cheap in enterprise software.
Contracts aren't.