Defog
3 known investors
Fine-tuned LLMs for enterprise analytics
Also known as Defog · Defog.ai
Investors · 3
Also in the syndicate · 1
Company profile
researched Aug 2026Defog (defog.ai) develops large language models and an AI data analyst product that let enterprise employees query structured data in natural language instead of waiting on analysts or building dashboards. Its core technology is SQLCoder, an open-source family of models for converting natural-language questions into SQL queries, which the company positions as the basis of its commercial offering. Users can ask questions of data, drill into specifics, explore hypotheses and ask follow-up questions, with results returned as tables, charts and reports.
The product connects to structured databases ranging from Postgres to Snowflake, to third-party tools via connectors, and to CSV files where no database exists. Defog's open-core models can be deployed and fine-tuned on-premises, and the company states that customer data is not shared with anyone, including its own model, since the model requires only database schema metadata. Defog also offers Defog Agents, which interprets user intent, explores hypotheses, writes code to extract data and run analyses, and returns results; the company says analyses that previously took months are completed in two to three days.
An earlier positioning of the product was an embedded, Q&A-style analytics interface added to a SaaS application via iframe or JavaScript plugin, allowing an app's end users to query structured data (SQL databases, warehouses) and unstructured corpuses (such as call transcripts) in everyday language.
Founding story
Co-founders Medha Basu and Rishabh Srivastava started Defog to address the difficulty of building good insights features into software products: teams focused on their core product and shipped limited dashboards, so Defog offered an embeddable natural-language Q&A layer for data. Srivastava is an ML developer who previously bootstrapped a data API company serving 100M+ monthly end users; Basu is a former journalist and editor who grew an enterprise content marketing firm to seven-figure annual revenue.
Business model
Open-core: the SQLCoder models are open source and freely downloadable, while Defog sells an enterprise offering built on those models that can be deployed and fine-tuned on a customer's own infrastructure, including on-premises.
Commercial enterprise offering layered on open-source models, sold to companies deploying Defog against their own databases.
Traction
SQLCoder models had been downloaded more than 50,000 times on Hugging Face and starred 1,800 times on GitHub within three months of release in 2023; the company later reported more than 300,000 downloads. Defog states its enterprise offering is used by healthcare, finance and analytics companies, and that a publicly listed US company deployed Defog to a 100-person department, cutting analysis time by 80% with estimated savings of 2,500 hours per month for its data science team.
Latest developments
Defog announced Defog Agents, which handles more complex "why" questions by inferring intent, exploring hypotheses, writing code to extract data and running analyses. The founders, Rishabh Srivastava and Medha Basu, are listed as active founders of FactIQ, which supplies clean, standardized investment research data (company financials, market prices, economic series and satellite imagery, with 100M+ data series covering the US and Asia) for AI agents.
▸Full profile — market position, technology, go-to-market, geography, history
Market position
Defog positions SQLCoder as the leading open-source model for natural-language-to-SQL conversion, comparing it directly against generalist models such as GPT-4. Hugging Face CEO Clement Delangue is quoted saying the best open source model for SQL is SQLCoder-34B, and the company cites endorsements from practitioners at AWS and JLL. VentureBeat included Defog in a January 2024 piece on how enterprises use open source LLMs.
Task-specialized open-source models that the company reports beat generalist GPT-4 models on text-to-SQL benchmarks, combined with on-prem deployment and fine-tuning, schema-only access for privacy, an open evaluation framework (SQLEval), and interpretability of results.
Technology
Defog's models are specialized LLMs for querying structured data. Its SQLCoder-34B model is reported by the company to outperform gpt-4-turbo and gpt-4 on the sql-eval benchmark for out-of-training-set schemas, and to reach 99%+ accuracy when fine-tuned to an enterprise's database schema. The company also maintains SQLEval, an extensible open-source evaluation framework for measuring model accuracy, and states its models are designed to be interpretable and explainable. Models can be deployed and fine-tuned on-prem, and require only schema metadata rather than customer data.
Go-to-market
Direct enterprise sales via contact forms and demo bookings, supported by open-source distribution of SQLCoder models on Hugging Face and GitHub, published benchmarks, and outreach through Twitter and LinkedIn.
Enterprises with structured data and analytics bottlenecks, including healthcare, finance and analytics companies, as well as data scientists and business leaders at large organizations; earlier positioning targeted SaaS founders and product managers wanting embedded analytics for their own users.
Geography
The founders' Y Combinator company profile lists Singapore as its location; customers cited include a publicly listed US company.
History
Defog was co-founded by Medha Basu and Rishabh Srivastava and went through Y Combinator's Winter 2023 batch, launching initially as an embedded "ChatGPT for data" analytics layer for SaaS applications. It open sourced its SQLCoder text-to-SQL models in 2023 and announced a $2.2M funding round in November 2023 led by Script Capital and Y Combinator. The founders subsequently appear on Y Combinator's company profile for FactIQ, a company providing investment research data for AI agents, based in Singapore.
Compiled by commissioned research from 5 cited public sources — announcements, filings, and press listed under research sources below.
Key figures
latest reportedCompany-reported or press-reported figures, each dated to when it was claimed — not independently audited.
Timeline · 5
launches, deals, and filingsListed in VentureBeat's "How enterprises are using open source LLMs: 16 examples".
Defog announced a $2.2M funding round led by Script Capital and Y Combinator, with participation from Hike Ventures, Pioneer Fund and angel investors including Dharmesh Shah and Divya Bhat, to fund open-source model research and productionization.
$2.2M source ↗
Defog introduced Agents, which interprets user intent, explores hypotheses, writes code to extract data, runs analyses and returns tables, charts and reports.
Defog open sourced SQLCoder, its family of LLMs for converting natural language questions into SQL; the 34B model is reported to beat gpt-4-turbo and gpt-4 on the sql-eval benchmark for out-of-training-set schemas.
Defog launched via Y Combinator as an embeddable, ChatGPT-style natural-language analytics interface added to SaaS apps by iframe or JavaScript plugin, querying structured databases and unstructured corpuses using schema metadata only.
Dated company events from announcements, filings, and press; legal rows summarize public dockets and regulator releases.
In the news
▸Research sources · 5
primary sources listed
- Defogdefog.com · web
5 public sources were cited for this profile; the first-party ones are listed here.
Frequently asked questions
- What does Defog do?
- Defog builds fine-tuned, open-core LLMs, including the SQLCoder text-to-SQL models, for enterprise data analysis.
- Who are Defog's investors?
- Defog's investors include Script Capital, Hike Ventures.