Text-to-SQL Analytics

The goal was simple: let someone ask a question in plain English and get a useful chart back. No SQL required.

I designed and implemented a Text-to-SQL interface inside an analytics SDK, translating natural language into safe, efficient SQL and a sensible visualization.

Problem

Business users struggle with SQL and most dashboard builders are powerful-but-clunky. The result is a constant dependency on data teams for simple questions.

Constraints

• Ambiguous queries are the norm (“revenue” can mean five different things)
• Schema size and business rules make naive prompting unreliable
• Latency matters (a 4–5s answer feels broken)
• Outputs need guardrails (wrong SQL is worse than no SQL)

Key decisions

• Schema-aware prompting (only relevant tables/columns, plus applicable business rules)
• Query generation with validation and predictable fallbacks for ambiguity
• Type-safe handling of model responses in TypeScript
• Visualization selection that prefers boring-but-correct defaults

At a high level the pipeline looked like:

1. Intent parsing - what the user wants to see
2. Schema linking - mapping domain language to columns/tables
3. SQL generation - optimized PostgreSQL queries
4. Chart selection - choose the right visualization and formatting

Example

User: "Show me monthly revenue growth for the last year"

Generated:
- SQL query joining revenue tables
- Line chart with month-over-month comparison
- Automatic formatting of currency values

Outcome

Enabled non-technical users to explore data independently, reducing the burden on data teams and accelerating insight discovery.