Hi HN, I built a fully local, execution-aware multi-agent Text-to-SQL system as a learning/research project.

It uses schema grounding (graph-based), multi-model SQL generation, execution sandboxing, and semantic validation instead of relying on prompting alone.

Evaluated honestly on Spider: • ~55% accuracy on single DB • ~34% cross-DB zero-shot

No APIs, no fine-tuning, everything runs locally.

Would appreciate feedback on the architecture and evaluation.