Project case study
DataLens
Schema-aware natural language to SQL. Introspects the database, validates the generated query, runs on read-only replicas, and auto-generates charts.
Private implementation; details are shared at an architectural level.
Problem
- Operational teams needed data but could not write SQL
- Existing BI tools required training nobody had bandwidth for
- Data team was drowning in ad-hoc report requests
Constraints
- Has to introspect the schema to avoid invalid queries
- No destructive operations, ever
- Ambiguous questions must ask for clarification, not guess
- Results need usable visualisations, not raw tables
Architecture
- Agent introspects the schema, generates SQL, and validates it against schema constraints
- All queries run on read-only replicas
- Visualisations auto-generated based on the shape of the result
Tradeoffs
- Graph-based reasoning over straight text-to-SQL. Introspection, validation, execution, and visualisation are distinct stages that needed explicit state
- Read-only replicas over prompt-only guardrails. Cannot trust a model to never generate a DELETE, however well the prompt is written
- Auto-visualisation over raw tables. Results need to be readable to non-technical users
Outcome
- Ops teams now self-serve most of their data questions
- Data team backlog dropped noticeably