All projects
LLM + Data
QueryWise
Governed text-to-SQL analytics assistant
Plain-English business questions become verified SQL, a table, a chart and a summary whose numbers are checked against the result — over a 45k-row e-commerce warehouse.
PythonsqlglotPostgreSQLpgvectorGeminiFastAPIDockerpytest
57/57
malicious queries blocked
0
false blocks
5 s
hard query timeout
274
automated tests
The problem
An LLM writing SQL can delete data, read personal information, or launch a query that freezes the database — and its summaries can quote numbers that are not in the result.
Architecture
- Question
- Schema linking
- LLM SQL draft
- AST guard
- Cost check
- Read-only execution
- Verified summary
How it works
- Semantic layer of approved metric definitions, so “revenue” always means the same thing.
- Embedding-based schema linking and few-shot examples ground generation in the real schema.
- Bounded self-repair (at most two attempts) and clarifying questions for ambiguous requests.
- Every number in the summary is verified against the query result.
Safety & reliability
- SQL AST guard rejects anything but a single read-only SELECT.
- Read-only database role limited to curated views without personal data.
- EXPLAIN cost ceiling and a 5-second statement timeout stop runaway queries.
What I learned
Most of the safety lives outside the model: a parser, a database role and a cost check are cheaper and more reliable than asking the model to behave.