This is an internal side project of ours that we decided to open source.
Besides being good at AE, it's pretty useful if you want to avoid huge data warehouse bills and your context window evaporating due to a giant query result.
Original author here, thanks for sharing.
Did anyone try to reproduce the results w/ reasoning models?
Very curious to see this.
A general meta-point: an LLM w/ no code generation and/or tool-calls will inherit non-trivial biases from its pre-training, post-training and safety guardrails