Who Is the Founder of Context AI?

Contextual AI (often searched as "Context AI") was co-founded in early 2023 by Douwe Kiela, who became CEO, and Amanpreet Singh, who became CTO. Both had worked as AI researchers at Facebook AI Research (FAIR, part of Meta) and at Hugging Face. Kiela is a natural language processing researcher and a co-author of the 2020 paper that introduced retrieval-augmented generation (RAG). The company builds enterprise AI systems that ground language model answers in an organization's own documents and data rather than in generic training data.

Several companies and products use similar names, so this article covers one of them: Contextual AI, the company at contextual.ai. Its name reflects its core technical bet. General-purpose models trained on internet text produce plausible answers, but enterprises need answers grounded in their own documents, policies, and data. Much of Kiela's research dealt with exactly that problem: how to connect language models to knowledge beyond their training text.

Douwe Kiela: Background and Research

Kiela earned an MPhil and a PhD in Computer Science at the University of Cambridge. His thesis, Deep embodiment: grounding semantics in perceptual modalities, worked on multimodal semantics: models that represent word meaning partly through perceptual information such as images, not only through other words. Grounding language in something outside the text runs through his later work as well.

At Facebook AI Research, Kiela worked on multimodal and grounded language learning and on how AI models are evaluated. Two contributions stand out. He is a co-author of Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020), the paper that introduced RAG, and Contextual AI's Series A announcement describes him as one of the pioneers of RAG. He is also the lead author of Dynabench, an open-source platform where people write examples that a model gets wrong but other people would answer correctly, so that dataset creation and model evaluation inform each other instead of relying on static benchmarks. Amanpreet Singh is a co-author of the Dynabench paper too.

After FAIR, Kiela was Head of Research at Hugging Face. He is also an adjunct professor in the Symbolic Systems Program at Stanford University.

These research threads share a common idea: language models perform better when they can draw on context beyond their training text. That idea became Contextual AI's founding premise.

Why Kiela Founded Contextual AI

By early 2023, large language models had demonstrated impressive general capabilities: writing code, summarizing documents, answering factual questions. But putting them to work inside enterprises ran into specific, technical obstacles.

LLMs hallucinated domain-specific answers. A model trained on internet text could produce confident-sounding but wrong answers about a company's internal policies, financial metrics, or product specifications, with no built-in way to say "I don't know this, let me check."

LLMs had no notion of access controls. In an enterprise, different users should see different information based on their role. A general-purpose model on its own has no concept of permissions, data classification, or need-to-know boundaries.

LLMs could not reach company knowledge on their own. Documents in SharePoint, data in Snowflake, policies in Confluence: a model needs a retrieval step to bring the specific context a user's question requires into the prompt.

RAG addresses these problems by retrieving relevant context from organizational sources first and then generating a response anchored in it. Contextual AI's argument, set out in its RAG 2.0 announcement, was that earlier RAG systems stitch together frozen models, vector databases, and poor-quality embeddings, and that the whole pipeline should instead be optimized end to end.

Contextual AI: Company Profile

Contextual AI came out of stealth in June 2023. According to its Series A announcement, most of the team works in office in Mountain View, California, with a presence in New York and London.

Seed (June 2023): $20M led by Bain Capital Ventures, with participation from Lightspeed, Greycroft, and SV Angel, announced in the company's launch press release.

Series A (August 2024): $80M led by Greycroft, with existing investors including Bain Capital Ventures and Lightspeed, and new investors Bezos Expeditions, NVentures (NVIDIA's venture arm), HSBC Ventures, and Snowflake Ventures. The announcement named two customers: Qualcomm, which had signed a multi-year contract, and HSBC, which planned to use the platform for AI-assisted knowledge management.

The two rounds total $100M. The company describes its product as a platform that lets enterprises build AI agents and applications that work securely on their own data.

In May 2026, Google DeepMind licensed Contextual AI's technology and hired part of its team, including Kiela, according to Bloomberg. As of 2026, Kiela's personal website lists him as a researcher at Google DeepMind and describes his Contextual AI role in the past tense, and Singh's personal website likewise describes his CTO role there as a previous position.

Technical Approach: RAG 2.0

Standard RAG follows a two-step process: a retriever searches a document store, selects relevant passages, and passes them to a language model that generates a response conditioned on the retrieved text. The retriever and the generator are separate systems, each built and tuned independently.

Contextual AI's approach, which the company calls RAG 2.0, optimizes the retriever and the language model together. Training backpropagates through both components, so the retriever is tuned toward passages the generator can use and the generator is tuned toward using retrieved evidence well. The goal is a pipeline tuned end to end for one objective: accurate answers grounded in the retrieved sources.

The argument is about accuracy. When retrieval surfaces irrelevant passages, a generator can fold them into a confident-sounding but wrong answer. Optimizing the components together is meant to make the retriever surface what the generator actually needs and to keep answers grounded in the retrieved evidence.

Context AI and Enterprise Data Governance

Enterprise RAG works only when the knowledge base it retrieves from is organized, governed, and trustworthy. This is the dependency that many RAG discussions skip over.

If a retriever searches a messy, undocumented data lake, the "context" it grounds on is unreliable. A document with an outdated product spec, a spreadsheet with manually adjusted revenue numbers, or a policy PDF that was superseded two quarters ago can all become sources of grounded-but-wrong answers. Without metadata, a retriever has no reliable way to tell current from outdated, canonical from draft, or governed from uncontrolled.

This is where data catalogs, business glossaries, and metadata management become prerequisites, not extras, for enterprise AI deployment. The knowledge base needs the same infrastructure that any data-driven organization needs: clear ownership, documented definitions, and recorded quality checks.

Without data governance, enterprise RAG amplifies the same metadata problems that plague traditional BI, except that the wrong answer now comes wrapped in a confident natural language response instead of a suspicious-looking spreadsheet cell.

Where Dawiso Fits

Dawiso provides the kind of governed metadata layer that enterprise RAG systems depend on. The data catalog documents what data exists and who owns it. The business glossary records terms, indicators, and rules, so that a question about "net revenue" can be matched to the definition the organization has agreed on.

Through the Model Context Protocol (MCP), AI agents can query Dawiso programmatically. The Dawiso MCP server lets an agent search the catalog, read object details such as descriptions and glossary definitions, follow data lineage, and retrieve data samples. A RAG pipeline can use it to check a metric's documented definition and where its data comes from before relying on retrieved content. The MCP server has no tools for data freshness or usage, so those checks stay with the systems that measure them.

Conclusion

Douwe Kiela and Amanpreet Singh founded Contextual AI in 2023 around a clear thesis: enterprise AI needs grounded context, not just bigger models. Kiela's research on grounded language at Cambridge and Facebook AI Research, including the original RAG paper, informed the company's RAG 2.0 approach of optimizing retrieval and generation together. The company raised $100M across its seed and Series A rounds and named enterprise customers such as Qualcomm and HSBC. The broader lesson extends beyond any single company: AI systems that retrieve from ungoverned knowledge bases inherit the problems of those knowledge bases, however sophisticated the model is.

Sources

• Contextual AI, Contextual AI Raises $80M Series A to Scale Production-Grade LLMs for Enterprises, 1 August 2024.
• Contextual AI, launch and seed funding press release (Business Wire), 7 June 2023.
• Contextual AI, Introducing RAG 2.0, 19 March 2024.
• Contextual AI, Company.
• Bloomberg, Google Hires Staff From AI Startup Contextual In Licensing Deal, 19 May 2026.
• Lewis, P. et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, NeurIPS 2020.
• Kiela, D. et al., Dynabench: Rethinking Benchmarking in NLP, 2021.
• Kiela, D., Deep embodiment: grounding semantics in perceptual modalities, University of Cambridge Technical Report UCAM-CL-TR-899, 2017.
• Douwe Kiela, personal website.
• Amanpreet Singh, personal website.
• Stanford Profiles, Douwe Kiela.
• Dawiso Help, Dawiso Model Context Protocol (MCP).