Rohit Patel, Meta's AI director and a core player in the company's machine learning infrastructure, is leaving to pursue his startup Ndoc. The move marks another departure from Meta's AI ranks as Patel launches a direct challenge to Microsoft Office and Google Workspace with AI-powered productivity tools.
Patel spent eight years at Meta building the company's AI evaluation frameworks and reinforcement learning systems, work that shaped how the social media giant trained and tested its large language models. His departure follows a pattern of senior AI talent leaving big tech to launch their own ventures, particularly in the generative AI space where tooling and productivity applications remain contested territory.
Ndoc represents a significant bet on AI-native document and productivity software. The startup enters a crowded field dominated by Microsoft Office (3 billion users globally) and Google Workspace (well over 1 billion users), but targets a gap: modern, AI-first workflows that both incumbents have been retrofitting with AI features rather than building from scratch. This approach mirrors successful startups like Notion, which disrupted note-taking, and Figma, which rebuilt design collaboration as a web-native product.
The competitive landscape for AI productivity has grown fierce. Competitors include Tome (AI-native presentations), Julius.ai (AI data analysis), and dozens of startups building specialized AI document tools. But none have attempted to fully replace Office and Workspace with a cohesive suite. Patel's insider knowledge of how AI systems work, particularly reinforcement learning techniques used to improve model outputs, could give Ndoc an edge in building smarter automation and suggestion features.
Patel's departure reflects broader trends in AI talent mobility. Senior researchers and engineers from OpenAI, Google DeepMind, Anthropic, and Meta have increasingly spun out to launch focused AI products rather than pursuing fundamental research alone. The market has signaled it will fund productivity-layer companies even as foundation model development concentrates at a few labs. Patel's name recognition in AI circles and his track record scaling systems at Meta's infrastructure scale provide credibility with investors.
The timing of his exit matters. Meta has been aggressive with its open-source Llama models and AI tools, pushing the industry toward commodity large language models. That commoditization creates opportunity for builders focused on applications and user experience rather than model training itself. Patel's experience evaluating AI outputs gives him insight into what enterprise customers actually need from productivity software, beyond feature parity with Office.
Ndoc's path forward depends on several factors. First, distribution. Microsoft and Google control the operating systems, cloud services, and enterprise sales channels that make productivity software sticky. Ndoc will likely target freelancers, small teams, and specific workflows first before attempting enterprise penetration. Second, differentiation. Simply wrapping AI around documents won't succeed. The startup needs to demonstrate that AI-first design genuinely improves how people work, not just add chatbots to familiar interfaces. Third, unit economics. AI inference costs remain high, and building profitable productivity software that runs AI on every keystroke requires either breakthrough efficiency gains or premium pricing that enterprises accept.
Patel's move underscores that deep learning expertise now flows toward commercial products rather than concentrated entirely at research labs. His eight-year foundation at Meta, combined with early-stage startup execution, positions Ndoc as a credible entrant in the AI productivity wars.