End-to-end natural-language SQL generation for complex multi-table schemas.

Highlights

• QLoRA fine-tuning of encoder-decoder models on Spider.
• 82% execution accuracy on unseen multi-table schemas.
• Schema-aware RAG with FAISS.
• Dockerized inference behind asynchronous FastAPI.

Built as a full serving path (train → retrieve schema context → API), not only a training run.