The business situation
Product and operations teams needed faster access to data, but routine questions still depended on analysts to translate business language into SQL and visual summaries.
Back to professional experience
An evaluated NL-to-SQL workflow for business teams
PRISM (Parent Company of OYO)
Built a two-agent natural-language-to-SQL workflow that helps product and operations teams query business data, execute SQL, and inspect visual outputs without analyst translation.
- Role
- Product Manager III
- When
- November 2025 - Present
- Product area
- AI PM, Data Products, SQL
The business situation.
Make operational and product data easier for business teams to query without waiting on analyst translation.
Evaluation result
85%
Accuracy against a golden dataset of business queries
Teams using it
6
Running their own analysis instead of routing through analysts
Workflow design
Two agents
SQL generation separated from execution and visualization
Evaluated at 85% accuracy against a golden dataset of business queries, and now used by six teams. The query set, the underlying tables, and internal usage data are confidential.
The decision I owned
I separated SQL generation from execution and visualization, so the product could measure generated-query quality and give users a clearer path to inspect the answer.
How I made the workflow trustworthy
Discovery centered on repeated business queries, common SQL patterns, data-table ambiguity, and the moments where teams waited for analyst support. I built a golden dataset of those queries with known-good answers, and made accuracy against it the release gate rather than a review step afterwards.
What shipped
A two-agent architecture generates SQL, executes queries, and returns visual outputs, with quality measured against the golden dataset. I own deployment and release validation. In parallel, I automated PM workflows for research, marketing benchmarks, daily Teams reporting, and PRD creation.
What changed
The workflow reached 85% accuracy on the golden dataset and rolled out across six teams, moving day-to-day analysis ownership from analysts to the product and business owners who ask the questions.
What I would carry forward
For AI products, a useful answer is not enough. The evaluation loop, the known-good comparisons, and a legible path to inspect output are part of the product itself.