The release of OpenAI’s o3 model has stirred discussions within the tech community. While AI’s growing role in development sometimes sparks concern, this advancement highlights how technology can complement and elevate human effort rather than replace it.

This study aims to show that synthetic data, enhanced by a human-in-the-loop process, can significantly improve model performance in Text2SQL tasks. By combining outputs from a state-of-the-art LLM with human expertise, we created a dataset of Natural Language Questions (NLQs) paired with SQL queries to fine-tune a language model. This hybrid approach delivered notable performance gains. Our long-term goal is to apply this methodology to other coding and reasoning tasks.

The client needed a solution to provide context-specific assistance by consolidating help information from multiple sources, including API documentation and Stack Overflow threads, to improve the productivity and quality of developer work.