Defog is a data-analysis company founded in 2023 through Y Combinator's Winter 2023 batch, built by a team based in Singapore. The company focuses on one problem: converting plain-English questions into accurate SQL queries against a company's own database schema.
Its open-source SQLCoder models, released in multiple sizes from 7B to 70B parameters, are fine-tuned specifically for text-to-SQL generation and are benchmarked against GPT-4 on Defog's own sql-eval framework. Defog reports the models were trained on more than 20,000 human-curated question-and-schema pairs, and the open-source releases have logged close to 20,000 downloads on Hugging Face.
Defog serves enterprise data and analytics teams that want natural-language querying without sending schema or query data to a third-party model provider. Pricing is not published on the website; prospective customers sign up and speak with sales for a quote, with enterprise contracts including custom SLAs and data-processing agreements.
