Disruption rarely arrives from the direction an industry expects, and oftentimes, outside of the industry entirely.

BM25 from scratch, a retrieval-vs-reasoning failure decomposition, and a fine-tuned cross-encoder. Leakage audits throughout; every number ...

In this repository it was rebuilt as an operable ML system: the data was audited and repaired, the baseline was re-measured, data and model were iterated on with tracked experiments, and versioned ...