Text-to-SQL and Data Interfaces
Turning questions in plain language into correct queries over real schemas, and knowing when the answer cannot be trusted.
Most data in an organisation is locked behind a query language that few of its people write. Text-to-SQL promises to remove that barrier, but a model that produces plausible SQL is not enough: the query has to be correct against a schema it has never seen, it has to run efficiently, and the system has to recognise when a question is ambiguous or unanswerable.
We work on the parts of that problem where data, not model size, is the bottleneck: schema representation and linking, synthesising training data that reflects real schemas, execution-grounded evaluation, and calibration so a system can say “I am not sure” rather than return a wrong table.