Production AI
AI text-to-SQL developer tool
A production tool that turns natural-language requests into optimized SQL, cutting analyst investigation time from about 25 minutes to under 8.
84% first-pass success120–180 queries/day · ~25 min → under 8 min
Problem
Analyst investigation time was about 25 minutes per query.
Approach
- Production text-to-SQL tool on OpenAI GPT-4o/o1. Ingested multi-source S3 data, applied AI preprocessing, used schema metadata + RAG for query generation.
- Python agentic query product translating natural-language requests into optimized SQL via semantic parsing and a schema-aware metadata catalog.
- Pipeline: ingestion, quality checks, metadata-aware text-to-SQL, validation/retry, cited responses.
- Evaluated LLM techniques: structured prompting, few-shot examples, continuous model evaluation.
My role
Owned the end-to-end pipeline.
Results
- Usage
- 120–180
- queries per day
- Accuracy
- 84%
- first-pass success
- Time
- 25 → <8
- minutes per investigation
Stack
- Python
- OpenAI GPT-4o/o1
- RAG
- AWS S3
- SQL