Ask a question in English. Get the right rows back.
A full stack platform with real-time dashboards and a Groq-backed assistant that turns natural-language questions into SQL.
- Role
- Everything
- Type
- Independent project
- Infra
- AWS, self-provisioned
- Repo
- github.com/abhijitsp11
- Year
- Personal
A side project about the gap between a question and a query.
Project data is relational, but the questions people ask about it are not. 'Which tasks slipped this sprint and who owns them' is one sentence and a three-table join.
Aster is a project management tool with real-time dashboards and an assistant that closes that gap — you ask in English, it writes the SQL, runs it, and shows you the rows.
The things that made it hard.
Schema context in the prompt does more for accuracy than a larger model does.
Generated SQL has to be constrained by the permissions of the person asking, not the service.
A wrong answer delivered confidently is worse than no answer.
How do you let a model near your database without letting it loose?
A model that writes SQL against production is a model that can write anything against production. Accuracy matters, but containment matters more.
2 passes at it.
Natural language to SQL
- Groq-backed assistant converts questions to SQL against the live schema.
- Reached 92% accuracy on the question set used for evaluation.
Infrastructure that stays up
- Provisioned production AWS from scratch — VPC, EC2, RDS and Cognito.
- RBAC applied at the data layer, so generated queries inherit the caller's scope.
What came out of it.
Built end to end — schema, API, infrastructure and interface — which is the part that made it worth doing.