Vivek Muppalla is vice president of AI engineering at Hippocratic AI, building clinical voice agents that help healthcare providers reach more patients while preserving safety, responsiveness, and human oversight. His career spans Amazon Photos, computer-vision simulation at Unity, synthetic data at Scale AI, enterprise language models at Cohere, and patient-facing healthcare systems.
At Scale AI, Muppalla became director of synthetic systems, working on training-data availability, privacy, and bias. His announcement about joining Scale connected synthetic data to obstacles slowing practical AI adoption, while acknowledging that generated datasets still require careful human judgment.
He subsequently became Cohere’s director of forward-deployed engineering and coauthored guidance on enterprise generative-AI projects. His work addressed retrieval quality, customized models, multilingual applications, private deployment, and enterprise data governance. He also contributed to the collective Command A technical report and highlighted Japanese enterprise-model development around Takane.
At Hippocratic AI, Muppalla applies these production disciplines to clinical conversations, where misunderstanding a medication, delaying a response, or scheduling the wrong appointment can directly affect patient care. His systems support proactive outreach and escalate concerning symptoms to human clinicians.
- Enterprise retrieval that reduces operating costs. Muppalla combines embeddings with cross-encoder reranking to select relevant documents before they enter a model’s context. His explanation of enterprise retrieval illustrates how sharper selection improves answer quality, reduces input-token costs, and supports citations and customer-controlled deployment.
- Polaris safety architecture. Hippocratic AI’s clinical system pairs a central conversational model with approximately 30 specialists covering medications, laboratory results, scheduling, and related tasks. Specialists operate in parallel, intervene selectively, and are supplemented by tool-call verification and escalation pathways to clinicians.
- Context-aware clinical speech recognition. Patient conversations involve noisy environments, short answers, unfamiliar drug names, and multiple languages. Muppalla’s approach supplements audio with conversation history and clinical context, uses an adapted Whisper component, preserves vocal cues, and applies additional checks to ambiguous single-word responses.
- Latency as a budget for clinical intelligence. Quantization, speculative decoding, and key-value-cache optimization create capacity for additional specialist reasoning and verification without sacrificing conversational responsiveness. His account of clinical voice-agent engineering emphasizes tail latency, clinician-assisted evaluation, and healthcare-specific failure modes.
- Empathy as a measurable safety requirement. The HEART benchmark for supportive dialogue assesses qualities including emotional responsiveness, attunement, and appropriate follow-through. Muppalla treats these interpersonal requirements as essential alongside clinical accuracy, speed, and human oversight.
