Project Two
A fintech startup had an idea for using natural language to let business users query their own financial data without writing SQL or relying on a reporting team.
The challenge
They had tried off-the-shelf tools and custom scripts but had not found something that was accurate enough, fast enough, and safe enough for production use. They needed a partner who could move quickly from concept to a usable product while keeping data governance and security front and centre.
Our approach
We ran a short discovery to validate the core assumption: that a combination of semantic search over metadata and a small language model fine-tuned on their schema could answer a high percentage of natural language questions with correct SQL and safe execution.
We made sensible choices for the first version:
- Clear separation between the NL-to-SQL layer and the execution layer
- Strict read-only access for the generated queries
- Human-in-the-loop step for any query that touched sensitive tables
We shipped at the end of week six with documentation, tests, and a simple UI.
Results
- Product now in use across multiple business units
- Client adopted the same sprint-based, hypothesis-driven approach for other data and AI initiatives
- User feedback from the pilot drove improvements to handling of ambiguous questions and better explanations for generated SQL
The engagement reinforced our view that discovery and delivery should be tightly coupled. For organisations that want to move fast without sacrificing rigour, this combination of aligned strategy, iterative delivery, and strong engineering practice remains the most reliable path to impact.
