Production-grade Asymmetric Multi-Agent Text-to-SQL on BIRD-SQL. Offline CHESS/FAISS pruning + MCI-SQL enrichment feed ≤3 Groq API calls: gpt-oss-120b generator · llama-4-scout reflector · gpt-oss-20b critic. LangGraph · SQLAlchemy sandbox · LangSmith · Docker · MIT.

Code for Struct-SQL paper

A privacy-preserving text-to-SQL workspace. Open-weight models on your machine, evaluated against BIRD.

Draft SQL and multi-candidate Text-to-SQL pipeline with schema compression, repair, and selector

Portfolio NL→SQL assistant: 94.0% execution accuracy on BIRD Mini-Dev (above the 92.96% human-expert baseline), schema-RAG + sqlglot AST guards + deterministic chart picker, $0 free-tier LLM stack. Streamlit chat UI, live demo. · Портфолио NL→SQL: 94% EA на BIRD Mini-Dev, schema-RAG + sqlglot-гарды + chart picker, бесплатный стек, живое демо.

A natural language to SQL interface powered by LLMs — ask questions in plain English, get database answers. Built on the BIRD-INTERACT benchmark with a Business Semantics Layer for accurate, context-aware query generation.

Reproducibility artifact for the paper 'Does Metadata Completeness Predict Text-to-SQL Accuracy? A Six-Model Ablation on BIRD Mini-Dev'.

Agentic Text-to-SQL trained with RL (GRPO): +7pt execution accuracy on full BIRD-dev from a general Qwen3.5-9B base