SolidLao/GenDB

Orchestrates five specialized LLM agents (Workload Analyzer, Storage Designer, Query Planner, Code Generator, Optimizer) to generate instance-optimized C++ execution code tailored to specific data distributions, hardware specs, and query patterns—avoiding the extensibility and economics constraints of general-purpose query engines. Achieves 2.8–5.0x speedups over DuckDB and Umbra by exploiting exact selectivities and hardware characteristics, with code verified through manual inspection and amortized generation cost across recurring workloads. Supports multi-model LLM providers (Claude, OpenAI Codex) and uses SQL template extraction for code reuse across parameterized queries.