# 200 Million Patient Interactions Later Vivek Muppalla, Hippocratic AI | AI Engineer World's Fair 2026 | 20:40 Source: https://www.youtube.com/watch?v=AN65uc645mE Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/200-million-patient-interactions-later Published: 2026-08-19 Tags: agents, evals, guardrails, multi-agent, voice ## TL;DR - Hippocratic AI uses AI care managers to make proactive clinical calls that clinicians cannot reach because of limited time and staffing. - Polaris runs 31 models in parallel, with one conversation model and 30 specialist models, to combine clinical coverage with phone-call latency. - The company evaluates safety through trained clinicians, synthetic data, harm-based grading, and a benchmark for empathy. ## Summary Vivek Muppalla argues that healthcare has been shaped by scarcity. Since clinicians do not have enough time to call everyone, systems triage patients and focus attention on the sickest cases. Hippocratic AI wants lower-cost clinical conversations to make proactive calls possible for more patients, while supporting clinicians rather than replacing them. He demonstrates an AI care manager that recognizes medications, clarifies vital signs, detects a medication issue, and escalates shortness of breath to a nurse. The engineering problem is combining clinical accuracy with low latency. Hippocratic AI built its own vertically integrated voice stack, including speech recognition that uses conversation context, domain knowledge, and prosody. Its Polaris architecture runs 31 models for each conversation, with specialists that speak only when they have relevant input. The company also uses tool-call verifiers, inference optimizations, clinician evaluations, and harm-based grading. Muppalla is direct about the standard required: a 1% scheduling error rate could send 100 patients a day to the wrong appointment type. ## Key ideas ### Lower conversation costs could let healthcare call every patient [00:01](https://www.youtube.com/watch?v=AN65uc645mE&t=1s) Muppalla begins with a question: almost nobody in the room had received a proactive call from a healthcare provider. He connects that absence to scarcity, with too few clinicians, too little time, and too little money. Healthcare therefore triages patients and gives attention to those who appear sickest. He says the arithmetic is changing because AI systems can now conduct clinically safe conversations while their cost continues to fall. The goal is to call everyone rather than reserve outreach for the sickest 5%. He frames the product as support for clinicians, helping them reach patients they could not reach on their own. ### The company builds around an employee oath to do no harm [01:54](https://www.youtube.com/watch?v=AN65uc645mE&t=114s) Hippocratic AI's stated principles are "do no harm," "patient first," and "access for all." Muppalla says these words are displayed across the company, but they are also an oath each employee takes on their first day. He says the product is built around that oath. At the time of the talk, he reports more than 200 million clinical interactions, deployment across more than 60 health systems, and an 8.5 out of 10 patient satisfaction rating. He also reports zero significant safety incidents. These claims set the standard for the architecture and evaluation work that follows. ### The care manager combines routine clinical work with escalation [02:44](https://www.youtube.com/watch?v=AN65uc645mE&t=164s) The product demonstration shows an AI care manager calling to check on a patient's health-monitoring devices. The agent identifies itself as AI when asked, then continues the conversation. It asks the patient to spell a medication name, confirms a blood pressure reading of 121 over 45 and a heart rate of 91, and asks whether a medication was stopped as instructed. When the patient reports weak legs and shortness of breath, the agent asks follow-up questions and recommends immediate conversation with a nurse. Muppalla uses the example to show that the system must handle recognition, clarification, medication questions, and escalation in one call. ### A phone system has to be both clinically accurate and fast [05:12](https://www.youtube.com/watch?v=AN65uc645mE&t=312s) Muppalla reduces the voice-system problem to intelligence and latency. Some clinically capable models take tens of seconds to answer or sometimes more than a minute, which does not work for a two-way telephone conversation. Faster models may not provide enough clinical accuracy for a safe call. Hippocratic AI wanted both properties, so it built a vertically integrated stack and optimized each part. Muppalla says the early versions also had several seconds of latency. After hundreds of optimizations, the product became much faster without losing its intelligence. The team benchmarks specific tasks such as lab-result checks and IVR navigation, where it wanted accuracy above 99%. ### Every latency gain is spent on more clinical reasoning [07:44](https://www.youtube.com/watch?v=AN65uc645mE&t=464s) Hippocratic AI treats latency improvements as capacity for more intelligence. When an optimization saves time, the team uses that time to add more reasoning or safety checks instead of simply keeping the faster response. Muppalla describes this as a flywheel: optimization creates a latency gap, the gap allows more intelligence, and the added system then creates new opportunities for optimization. Polaris, the company's constellation architecture, has a hearing system on the left, a reasoning system in the middle, and a speaking system on the right. The whole round trip has to remain fast enough for live conversation. ### Polaris uses specialist models to reduce single-model failure [08:36](https://www.youtube.com/watch?v=AN65uc645mE&t=516s) Polaris runs 31 models for each conversation. One central model manages the conversation, while 30 specialist models cover areas such as labs, medications, and scheduling. The specialists run in parallel, but each first makes a quick decision about whether it has anything to add. If it does not, it short-circuits, which keeps the system within its latency budget. Relevant specialist input is passed to the central model, which uses it with the rest of the conversation and the overall task. Muppalla says the architecture provides redundancy because a single model would be a single point of failure in a patient conversation. ### Speech recognition needs clinical context and a second pass for one-word answers [09:29](https://www.youtube.com/watch?v=AN65uc645mE&t=569s) Muppalla says real-world calls are noisy, while many speech benchmarks use quiet recordings. Apparent reasoning failures often begin as recognition failures. Hippocratic AI's decoder-only audio language model receives the audio, the conversation context, and domain knowledge. Its encoder is based on an open-source Whisper v3 large turbo model fine-tuned on millions of clinical conversations. A conformer projector compresses the audio while preserving pauses and stress, so the model hears how something was said as well as what was said. Medication context narrows recognition to a finite list. Single-word answers receive a second scoring pass because "a now" can become "a no," which can change a clinical decision. ### Safety evaluation has to account for failures at production scale [16:25](https://www.youtube.com/watch?v=AN65uc645mE&t=985s) Muppalla says 99% accuracy is inadequate for scheduling. With more than 10,000 calls a day, a 1% failure rate could give 100 people the wrong appointment type, causing them to arrive at the wrong date or time or miss a critical appointment. He says roughly 450 tests are needed to be 99% sure of catching a 1% error rate, and about 1,900 tests to see that error occur 10 times. Hippocratic AI combines synthetic data with human review. More than 7,000 trained clinicians have evaluated close to 800,000 clinical conversations. The company grades outputs on correctness and harm, including minor harm, severe harm, and death. It also built the HEART benchmark to measure empathy. ## Notable quotes - "We're trying to build clinically safe abundance for all." (01:05) - "The brain isn't a singular model." (08:32) - "For us, even the 99% is pretty bad." (16:32) - "Safe is not enough, right? A key part of the product is the empathy." (18:40) ## Tools & references mentioned - Hippocratic AI - Polaris - Whisper v3 large turbo - HEART - OpenAI ## Who should watch - You are building voice agents for healthcare and need to handle clinical recognition, escalation, and live response times in one system. - Your agent uses tool calls or scheduling workflows where a small error can send patients to the wrong appointment. - You need evaluation methods that measure harm and empathy alongside ordinary task accuracy.