After eight years at Meta, Rohit Patel is leaving to try to replace Microsoft Office and Google Workspace. TechCrunch reported it exclusively today (October 27), and I checked the parts that can actually be checked.
The verified facts first. Patel was a director at Meta Superintelligence Labs, working on reinforcement learning and on evaluating whether AI output is trustworthy. His new company is called Ndoc. It is a team of three, with no outside funding — Patel told TechCrunch he "didn't want to take money until I had some proof that people love the product." As of today it is running an enterprise pilot, opening an alpha to a limited waitlist, with a beta planned in the coming weeks and a full launch next year. Customers will pay for usage and tokens.
I also probed the site itself, because a pitch is not a product. ndoc.ai is live: it returns HTTP 200 with an 806-byte single-page-app shell, which loads a 786,891-byte JavaScript bundle. That bundle is not a placeholder — it contains the product's own vocabulary (Stage appears 31 times, Narrative 29 times, Model 10 times, .ndoc.html 4 times) and a real sign-in flow, including a POST /api/auth/sign-in path whose code treats a 202 response as "waitlisted." So the alpha waitlist described in the article exists in the shipped code, not just in the interview.
The one idea worth taking seriously
Ndoc is organized into three tabs. Stage is where the presentation lives — slide decks, dashboards, simulations, mini apps. Narrative is the analysis, with numbered citations and generated figures: the "why" behind the numbers. Model is where the underlying datasets and Python scripts live.
That third tab is the whole argument. Patel's background is in checking whether AI output deserves trust, and Ndoc applies it literally: if the AI built your spreadsheet or deck, you can open the Python script that produced a number and check it. If you don't read Python, you can ask the built-in chat which line contains your formula, and verify the formula from there. "There's nothing hidden behind the curtain here," he said.
This is the correct diagnosis of the current generation of AI office tools. Bolting a model onto a spreadsheet so it fills cells for you — Patel's line was that this is like discovering electricity and using it to put lights on the bulls that till your soil, instead of building a tractor — produces output that looks finished and is hard to audit. A generated number with no provenance is worse than no number, because it gets pasted into a decision. Attaching the generating script to every figure does not make the figure right, but it makes it falsifiable, and that is a different and much better failure mode.
The second idea is the file format. Documents export as a single .ndoc.html file that opens in any browser, and Patel says he plans to open-source the format so other developers can build compatible tools. He says that plan is part of why he left: getting agreement on an open format inside a company Meta's size would have meant "hundreds of people and opinions."
Where I am skeptical
Two things, and they are both in the reporting itself if you read it carefully.
First, "plans to open-source" is not "open-sourced." The .ndoc.html format is currently a promise from a three-person startup with no outside funding. Formats only break lock-in when other tools actually read and write them, and that requires a spec, a reference implementation, and time. Today, an exported .ndoc.html is portable in the weakest useful sense: a browser can open it. That is already better than a proprietary binary, but it is not yet an ecosystem.
Second, the competition is not standing still, and it owns the habits Ndoc needs to break. OpenAI, Anthropic, Google, Meta, and SpaceXAI are all shipping enterprise tools that work inside the Workspace and Microsoft 365 documents people already have, and OpenAI launched Space — a shared workspace inside ChatGPT — last month. Ndoc's answer is connectors to import existing data, which helps with files and does nothing for the harder problem: people do not switch the surface where their team's work lives because a new one is more principled.
Ndoc is also built on a mixture of open-source models, fine-tuned in-house, which Patel says makes it considerably cheaper to run than tools built on frontier models. That is a claim from the founder, not a number I can verify from outside — there is no pricing page with token rates to check yet, and alpha customers are the first real test of it.
My take
I would not bet on a three-person team displacing Office. I would bet on its central design rule spreading: an AI-generated document should ship with the code that generated its numbers, in the same file, where a reader can run it. That rule costs incumbents almost nothing to copy, which is exactly why a startup has to move first — inside a large company it dies in committee, as Patel learned.
If .ndoc.html does get open-sourced with a real spec, it becomes the interesting artifact here, more than the app itself: a document format where the narrative, the presentation, and the generating model are one self-contained, browser-readable file. That is a format a developer can build on without asking anyone's permission. Until that spec ships, Ndoc is a well-aimed alpha with a waitlist I verified is real — and a diagnosis of AI office software that I think is right, whoever ends up winning it.
Sources: TechCrunch, October 27, 2026; ndoc.ai live probe, October 27, 2026 (HTTP 200, 806-byte app shell, 786,891-byte bundle, sign-in/waitlist flow in the shipped code).
Top comments (0)