Vivek Muppalla presents Hippocratic AI's approach to healthcare voice agents, with the aim of extending clinician outreach. The talk's interaction volume, patient satisfaction and safety results are company-reported claims. The technical discussion explains how speech recognition, specialist reasoning, tool verification and inference optimization are combined under a tight response-time constraint.
Vivek Muppalla argues that apparent reasoning failures can originate in misheard audio, particularly with medication names, language switching and short replies. The described speech system combines an audio encoder and a projection step with conversation history and domain context. Additional scoring of brief responses is intended to catch ambiguities before the transcript reaches the reasoning system.
Vivek Muppalla describes a central conversation model supported by specialist models for tasks such as medications, lab information and scheduling. Specialists run in parallel and first check whether their input is relevant, allowing many to exit quickly. Separate verifiers inspect tool-call inputs and outputs, with some checks continuing asynchronously or after a call when the task allows correction.
Vivek Muppalla treats latency savings as capacity for additional reasoning and verification. The implementation examples include lower-precision computation, speculative decoding and KV-cache compression. These are presented as optimizations within the company's system; the talk does not independently establish that every optimization preserves quality or that the reported speed improvements generalize to other workloads.
Vivek Muppalla explains why even a small error rate can create many failed appointments at high call volumes. The evaluation process described combines synthetic cases with clinician-led conversations and grades potential harm as well as answer correctness. The talk also discusses evaluating empathy, but its comparative safety figures and clinical performance claims remain dependent on the company's evaluation methods and definitions.
Watch on YouTube
