Chaitanya Asawa describes a progression from ambient clinical documentation toward contextual decision support and related workflows. He explains that clinical notes affect both subsequent care and billing, making their accuracy consequential. The proposed systems combine electronic health record context, live conversations, medical literature and clinical guidelines.
Chaitanya Asawa presents evaluation as an ongoing engineering process: internal benchmarks precede staged clinician rollouts, followed by monitoring and feedback. He describes separate signals for clinical quality, safety, adversarial behavior, and product tone, rather than relying on a single score or a few successful demonstrations.
Chaitanya Asawa argues that verifying a clinical answer can be nearly as difficult as generating it. His team therefore uses real cases and physician-authored rubrics instead of demanding one exact reference answer. He describes independent rubric creation, adjudication by another physician and further clinical quality assurance, with language-model judges checking responses against the resulting criteria.
Chaitanya Asawa explains two approaches to cost and latency. Decomposing documentation into specific workflows allows smaller post-trained models to handle tasks such as individual note sections. For in-visit orders, cheaper event-detection gates trigger heavier models to match relevant approved orders, which are queued for clinician signoff in the electronic health record. The talk describes an engineering approach and company experience, without independently establishing clinical effectiveness.
Watch on YouTube
