Sreeram Nurani Ramasubramanian

Replacing end-to-end LLM reasoning with a hybrid extraction-and-rule architecture drove a 24.5-point jump in emergency recall while eliminating prompt regression testing across 150,000 real-world clinical queries.