Correct SQL from plain English, against a database built to resist it
of analyst queue reduced to seconds.
- 2 to 3 days to under 30 secondsin the data team's queue, self-serve
- About 92%execution accuracy on the correctness benchmark, up from roughly 60% for an off-the-shelf approach
- About a dozen to 200+from SQL-fluent analysts to business users who can ask the database directly
- Roughly 40%of the data team's week returned to platform work (estimated with the team)
// The problem
Where time went
The database was fast because it was unusual: a specialized SQL dialect, approximate-count sketches, unsupported joins. Most non-trivial questions went to a small data team and sat 2 to 3 days in the queue, engineers who should have been building the platform answered one-off queries, and off-the-shelf text-to-SQL returned confident wrong numbers with no error at all.
// The build
What we built with them
LevelUp built for correctness first, with the data team: schema-aware retrieval so the model cannot invent fields, query planning, dialect-correct generation with the sketch and time functions injected, static validation before execution, and execution-guided self-correction bounded to a few attempts. Approximate results are surfaced as estimates. Underneath it, a correctness benchmark scored on execution accuracy, the number almost no text-to-SQL project measures.
// Where it stands
In production with the data team; decisions run on current data instead of on whatever the queue delivered days later.
// How we measured
Execution accuracy on a domain gold set of questions with expected results, tracked with latency and cost; time to answer against the workflow-review baseline.
// Build with confidence
Talk to us about testing your agents before launch.
