Business#Model training#Office suite
Meta Superintelligence Labs director Rohit Patel leaves to build Ndoc, an AI-native office suite on fine-tuned open models
After eight years at Meta on RL and AI evaluations, Superintelligence Labs director Rohit Patel left to build Ndoc: three people, no funding, .ndoc.html export.
Rohit Patel, who spent eight years at Meta working on reinforcement learning and AI evaluations and left his post as a director at Meta Superintelligence Labs, has built Ndoc — an AI-native office suite, running on fine-tuned open models, whose documents export as a single .ndoc.html file that any browser can open. Per TechCrunch’s exclusive on October 27, the company is three people with no outside funding, is running enterprise pilots, and has an alpha limited to its waitlist; a beta is planned within weeks. The timing makes it a counterexample: while every major lab staples AI onto an existing office suite, Patel is rebuilding the document from its structure up.
Key points
- The person: Patel spent eight years at Meta on reinforcement learning and AI evaluations, leaving with the title of director at Meta Superintelligence Labs; Ndoc became public through TechCrunch’s exclusive on October 27.
- The product: Ndoc splits every document into three layers — Stage (presentations), Narrative (the analysis), Model (datasets and Python scripts) — so any AI-generated claim can be traced down to the script behind it, and the whole thing exports as one
.ndoc.htmlfile whose format Patel plans to open-source. - The economics: three people, zero funding, inference costs kept far below frontier-model rivals by fine-tuning open models; free sign-up, pay per usage, beta in weeks.
Background
The departure date matters. The day before Patel’s exit made news (October 26), Meta filed with the SEC for an equity raise of up to $95 billion aimed at AI infrastructure, and two weeks before that it open-sourced Muse Spark Open, an 816B-parameter model — betting on compute and models at the same time. Patel told TechCrunch that open source was part of why he left: at a company Meta’s size, open-sourcing a format takes “hundreds of people and opinions” to align, and he wants to “move at the pace of what AI can do.”
The market he is entering is not short of incumbents. Per market figures assembled by TechBuzz.ai, Microsoft 365 Office has more than 345 million paid seats and Google Workspace more than 9 million paying businesses — secondhand numbers we could not audit against primary filings, though they are consistent with what both companies have disclosed publicly. The incumbents are moving too: OpenAI shipped Dots, its always-on cloud agents, at the end of September, and Google launched its Gemini agent wired into Workspace on October 8. Both bolt agents onto existing suites; neither rebuilds the document.
The facts
Ndoc’s structure has three layers, per the company, each holding something you can inspect:
| Layer | Holds | What you can check |
|---|---|---|
| Stage | Slides, dashboards, simulations, mini apps | The presentation layer itself |
| Narrative | Analysis, numbered citations, generated charts | The reasoning behind the numbers |
| Model | Underlying datasets and Python scripts | Formulas and scripts open to inspection |
“There’s nothing hidden behind the curtain here,” Patel told TechCrunch — users can read the Python script that generated a document and verify its formulas. On format: every document exports as a single .ndoc.html file, openable in any browser, and Patel plans to open-source the format so other developers can build compatible apps. The model layer is “a mixture of open source models,” built and fine-tuned in-house; because the product was built from scratch, TechCrunch reports, running costs sit far below frontier-model equivalents. Memory persists across four scopes — personal, single document, team, and company — with audience and tone set as context.
As of October 27, availability is: free sign-up with usage- and token-based pricing, enterprise pilots underway, an alpha for waitlist users, a beta planned within weeks, and a general launch sometime next year. The ndoc.ai domain 302-redirects to alpha.ndoc.ai. We could not find an announcement post from Patel himself on X or LinkedIn; every product detail here traces to TechCrunch’s exclusive by Rebecca Bellan.
What others say
Two Patel quotes are worth keeping. On the competition: rivals are attaching AI to spreadsheets to auto-fill cells, and “this is like if you just figured out electricity, and then you decided the best use case is to put lights on my bulls that till my soil. No. Get a tractor.” On the risk: he concedes OpenAI’s Space could be a threat, but thinks the labs “largely have the wrong idea” — bolting AI onto existing tools “just sort of kills my soul.”
Third-party scrutiny is absent so far. As of October 28, searching Hacker News for “Rohit Patel” returns zero stories, and “ndoc” surfaces only unrelated old threads. In Chinese coverage, IT Home ran a write-up that inflates his title to the equivalent of “head” — the TechCrunch original says director, and we follow the original. TechBuzz.ai attributes its entire piece to TechCrunch and adds no independent sourcing, though it contributes the market figures above. One day after publication, no independent review or critique of Ndoc exists.
Our take
What makes Ndoc worth watching is the bet it places in the AI-native-versus-bolted-on debate: the incumbents staple AI onto suites that already own the market, while Patel is testing whether an AI-native document can grow from scratch. Three layers plus a single exportable HTML file turns “auditable AI output” into the file format itself — and if the .ndoc.html format really does go open source, it could become the interoperability base layer for this category, which would matter more than the suite.
The honest risk list is long too: three people, no funding, no beta yet. The claim that fine-tuned open models can match a frontier-model experience is, so far, his own claim — no benchmark backs it. And enterprise office moats have never lived in document formats; they live in Excel macros, org permissions, and decades of accumulated workflow. The shortest verification a reader can run today is joining the waitlist, then — when the beta lands in a few weeks — opening the Python script under an AI-generated document. Whether that layer is as transparent as described is the first hard test of whether this story holds.