--- schema: 1 type: video title: '200 Million Patient Interactions Later — Vivek Muppalla, Hippocratic AI' description: >- Asked how many in the room had ever received a proactive call from their healthcare provider, almost no hands went up. Vivek Muppalla treats that as the signature of scarcity: too few clinicians and too few hours, so the system triages and only the sickest get called. Hippocratic AI is past 200 million clinical conversations across more than 60 health systems, which is what changes the arithmetic. The tradeoff they refused is the familiar one: models accurate enough for this work can take tens of seconds to answer, and models fast enough for a phone call are not safe enough to make one. So the stack was built end to end. Polaris runs 31 models on every conversation, one holding the thread and 30 specialists covering labs, medications and scheduling, executed in parallel with each specialist first making a fast check on whether it has anything to say at all. A single model would be a single point of failure. Their speech recognition is a decoder only audio system fed the conversation so far and the domain context alongside the audio, so a drug name resolves against a finite list rather than an unbounded one, and prosody survives the projection so it hears the how as well as the what. Single word answers get a second scoring pass, because a heard as no is catastrophic. On evaluation he does the arithmetic that makes 99% unacceptable: at 10,000 calls a day that is 100 people sent to the wrong appointment, and catching a 1% failure rate takes roughly 450 tests. Speaker info: - https://x.com/vim1up - https://www.linkedin.com/in/vivekmuppalla/ - https://hippocraticai.com/ Timestamps: 0:00 - Who has ever been called by their provider 1:05 - Triage as a math problem, and what flips it 1:54 - The oath every employee takes 2:44 - A call with a patient, start to escalation 5:12 - Why a generic voice stack does not work 6:04 - Building vertically to get speed and intelligence 7:44 - Latency and intelligence as a compounding flywheel 8:36 - Polaris, and running 31 models at once 9:29 - What speech recognition gets wrong in the real world 10:27 - Feeding context and domain knowledge into the audio model 11:19 - A finite list of medications instead of an infinite one 12:10 - Rescoring single word answers 13:51 - How a specialist decides to speak up 14:43 - Verifiers for tool calls 15:34 - Quantization, speculative decoding, cache 16:25 - Why 99% is a bad number here 17:14 - The tests needed to catch a 1% failure 18:03 - Grading on the scale used for humans 18:55 - Building a benchmark for empathy canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/200-million-patient-interactions-later-vivek-muppal-901df2fcf854 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/200-million-patient-interactions-later-vivek-muppal-901df2fcf854.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T05:08:00.740Z' language: null children: [] videoId: AN65uc645mE durationSeconds: 1240 publishedAt: '2026-08-19T16:00:06.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 23 endSeconds: 70 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=23' heldReason: rights - ordinal: 1 text: null startSeconds: 23 endSeconds: 70 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=23' heldReason: rights - ordinal: 2 text: null startSeconds: 82 endSeconds: 114 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=82' heldReason: rights - ordinal: 3 text: null startSeconds: 179 endSeconds: 312 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=179' heldReason: rights - ordinal: 4 text: null startSeconds: 179 endSeconds: 312 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=179' heldReason: rights - ordinal: 5 text: null startSeconds: 322 endSeconds: 364 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=322' heldReason: rights - ordinal: 6 text: null startSeconds: 375 endSeconds: 471 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=375' heldReason: rights - ordinal: 7 text: null startSeconds: 482 endSeconds: 523 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=482' heldReason: rights - ordinal: 8 text: null startSeconds: 580 endSeconds: 636 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=580' heldReason: rights - ordinal: 9 text: null startSeconds: 636 endSeconds: 648 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=636' heldReason: rights - ordinal: 10 text: null startSeconds: 682 endSeconds: 692 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=682' heldReason: rights - ordinal: 11 text: null startSeconds: 738 endSeconds: 748 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=738' heldReason: rights - ordinal: 12 text: null startSeconds: 748 endSeconds: 837 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=748' heldReason: rights - ordinal: 13 text: null startSeconds: 849 endSeconds: 893 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=849' heldReason: rights - ordinal: 14 text: null startSeconds: 904 endSeconds: 938 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=904' heldReason: rights - ordinal: 15 text: null startSeconds: 938 endSeconds: 949 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=938' heldReason: rights - ordinal: 16 text: null startSeconds: 1001 endSeconds: 1035 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=1001' heldReason: rights - ordinal: 17 text: null startSeconds: 1035 endSeconds: 1047 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=1035' heldReason: rights - ordinal: 18 text: null startSeconds: 1047 endSeconds: 1091 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=1047' heldReason: rights - ordinal: 19 text: null startSeconds: 1091 endSeconds: 1105 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=1091' heldReason: rights - ordinal: 20 text: null startSeconds: 1141 endSeconds: 1153 watch: 'https://liquidgold.cc/watch/AN65uc645mE?t=1141' heldReason: rights --- # 200 Million Patient Interactions Later — Vivek Muppalla, Hippocratic AI Asked how many in the room had ever received a proactive call from their healthcare provider, almost no hands went up. Vivek Muppalla treats that as the signature of scarcity: too few clinicians and too few hours, so the system triages and only the sickest get called. Hippocratic AI is past 200 million clinical conversations across more than 60 health systems, which is what changes the arithmetic. The tradeoff they refused is the familiar one: models accurate enough for this work can take tens of seconds to answer, and models fast enough for a phone call are not safe enough to make one. So the stack was built end to end. Polaris runs 31 models on every conversation, one holding the thread and 30 specialists covering labs, medications and scheduling, executed in parallel with each specialist first making a fast check on whether it has anything to say at all. A single model would be a single point of failure. Their speech recognition is a decoder only audio system fed the conversation so far and the domain context alongside the audio, so a drug name resolves against a finite list rather than an unbounded one, and prosody survives the projection so it hears the how as well as the what. Single word answers get a second scoring pass, because a heard as no is catastrophic. On evaluation he does the arithmetic that makes 99% unacceptable: at 10,000 calls a day that is 100 people sent to the wrong appointment, and catching a 1% failure rate takes roughly 450 tests. Speaker info: - https://x.com/vim1up - https://www.linkedin.com/in/vivekmuppalla/ - https://hippocraticai.com/ Timestamps: 0:00 - Who has ever been called by their provider 1:05 - Triage as a math problem, and what flips it 1:54 - The oath every employee takes 2:44 - A call with a patient, start to escalation 5:12 - Why a generic voice stack does not work 6:04 - Building vertically to get speed and intelligence 7:44 - Latency and intelligence as a compounding flywheel 8:36 - Polaris, and running 31 models at once 9:29 - What speech recognition gets wrong in the real world 10:27 - Feeding context and domain knowledge into the audio model 11:19 - A finite list of medications instead of an infinite one 12:10 - Rescoring single word answers 13:51 - How a specialist decides to speak up 14:43 - Verifiers for tool calls 15:34 - Quantization, speculative decoding, cache 16:25 - Why 99% is a bad number here 17:14 - The tests needed to catch a 1% failure 18:03 - Grading on the scale used for humans 18:55 - Building a benchmark for empathy Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/200-million-patient-interactions-later-vivek-muppal-901df2fcf854 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.