Name: From 0 to 1: How Zingage Built AI-Powered Home Care on Claude
Uploaded: 2026-02-26T19:30:06.831Z
Duration: 1 h 28 min 33 s
Description: From 0 to 1: How Zingage Built AI-Powered Home Care on Claude

Transcript for "From 0 to 1: How Zingage Built AI-Powered Home Care on Claude": Hi. Hi, everyone. Welcome to today's conversation with Anthropic and Zingage. My name is Araba Koomson. I am on our start ups team here at Anthropic, and I am super, super excited to dive into how's Zingage built AI powered home care with Claude. I'm gonna run us through actually a few housekeeping notes before we begin. So first of all, if you are afraid that you will miss something from the webinar, do not fear. We have a recording of this webinar, and it'll be distributed through your emails within, like, twenty four hours. Second, if you have any questions, please don't wait until the end to ask them. You can actually just pop them in as they come to you through the submit a question widget that's on the right hand side of your screen. And then we'll either answer it as best as possible during our time, and then we'll also reserve some time for Q&A at the end as well. And then finally, we love feedback. Please, please, please either, like, rate the webinar by selecting the survey widget at the bottom of the screen at any point during the webinar, and then there'll be other opportunities to give feedback as well. As I mentioned, my name is Araba Koomson, and then it's a pleasure to be joined today by two incredible speakers. So we'll have Victor Hunt, who is the cofounder and CEO at Zingage, and then William Hu, who's one of our incredible technical AI experts here at Anthropic. A little bit more about the agenda for today. So first, we're gonna cover how AI is transforming home care operations, everything from scheduling to compliance. And we'll walk through Zingage story on how they build on Claude for more than 400 agencies, And that'll actually include a live demo of AI's role in the front office. And then finally, we'll close out with the Q&A that I mentioned. So Victor and William will come back on screen, and they'll share practical lessons for scaling agentic AI in health care. Before we jump into Zingage's impact, I actually wanted to take a quick look at how AI is transforming health care across time. So in less than, I'd say, about a decade, AI went from a research breakthrough to a technology that can actually reason, write code, even, like, take autonomous action. Health care has tracked at that same curve. You can see a lot of it in this graph here, even outpacing the curve sometimes. So for example, AI can now catch heart disease from routine EKGs. We saw some interesting studies on how it can reduce missed cancer screenings by ten percent and then also help elderly patients manage chronic conditions at home. Increasingly, AI isn't just like working in the background on data and scheduling and ops and analytics. Instead, it's really in the foreground. So calling caregivers, filling open shifts, and keeping operations running in real time. That's exactly why we're excited to have Zingage join us today and talk about what they've built. Victor's team built Zingage using what we call the Claude Thinking Engine. So it's like our full platform stack. At the foundation, frontier models are building to be agentic from the ground up. And then on top of that, they're building blocks for production agents. So you can think of that as tools, knowledge, and developer services. Then at that outcomes layer, so that's the block right at the top, you can use cloud directly either through out of the box products for employees and your teams. So I'm sure a lot of you have been following along our coworker announcement. You can think about coworker, chat, and code. Or at the outcomes layer, you can think about how you build your own agents either on our API or with the agents SDK, which is actually what's Zingage did. So to dive deeper into the story, I'll hand it off to Victor next. Thanks, Araba, and thanks, everybody, for joining today. I'm thrilled to introduce the work we're doing, talk about how we've partnered with Anthropic to make this work successful, and to answer questions for folks who are also building, in the AI space or health care as well. So just to start, as Araba mentioned, I'm Victor Hunt. I'm the CEO and cofounder here at Zingage, and we are building, the only end to end agent that solves care coordination for home based care. So everything from, phone answering, front office, scheduling, compliance, and all of the nuances of operations that make home based care possible from a logistical standpoint so that folks can get into the home, and safely receive the care they need. So before we jump into what we do, I wanna talk about why we do it. My inspiration for building Zingage starts with my family. I grew up, in a family of caregivers, and folks whose lives depended on health care. My biggest inspiration in my life is my grandmother, who worked the emergency room floor at Jamaica Hospital in Queens, New York for over thirty years, And her vision in becoming a nurse started with her family members who needed care and understanding how important it was to not only service folks at the most acute needs there in the emergency room, but to get ahead of it really in the places where they need care the most and where we can have the highest impact, which is in the home. And that inspired her to start her own home care business many years ago when things were entirely pen and paper. And like many nurses who follow that path, she quickly realized that, managing a fleet of nurses and all of the requirements around, the logistics of home care is a lot more difficult than being a clinician herself. And so being that her first love was in the health care and not so much the operations, she went back into that clinical practice as opposed to continuing that business. And I see this now all the time with the providers who we work with. The bait and switch of the promise of having such a widespread impact to later find how much more this has to do with being able to keep up with phone calls and keep up with regulation rather than delivering the great care that inspired you to get started. And so I want to actually pause for a second and ask the audience to take a quick poll that we're going to put up and, basically, to understand who among us has experienced home care in their lives, be it someone you know who has received care or is currently receiving care. And the reason I ask this is because I want to share what this looks like on the large on the large scale across the board and why this problem matters to us, more deeply than just my personal experience. And that points to, just overwhelming trends across the country and the world, frankly. We hear it all the time, the demographic shift, the gray tsunami, 11,000 seniors, turning 65 every day. And what's even more concerning is when we double click into this and see that, over twenty eight percent of folks 65 live alone, twenty five percent of folks with dementia living alone, and many of these people living far from the kind of care that you would expect if you were in those sort of conditions. And this is where home care comes in to, truly make a huge difference in meeting people where they are, and especially at those moments where that care can have, the most impact in preventing, those acute needs from becoming hospitalization or worse. And so to solve those problems, requires an immense amount of logistical, a logistical operation, from the care provider side, which is what we at Zingage serve. We provide our provider customers, those home care agencies, the infrastructure to automate all of the work behind the scenes that allows them to ensure care is delivered, expand access to care, and to stay in compliance, with the regulations that drive this industry. And so as folks answer that question, I'm going to shift over into a little bit more about what we do and how we deploy our agents in the context of home care. So at Zingage, our job is to solve for the logistical aspects of care coordination, and that is to the extent of where our providers need help the most. And so the first part of this is, for us, being able to answer and receive every inquiry, and that is entirely phone based. These folks are calling in all of the time to ask questions about things like schedules, to report emergencies, whether that be hospitalizations, last minute cancellations, or concerns such as abuse and neglect. And so that's where our first agent, Riley Receptionist, comes in. That is the 24/7 front desk for home based care, answering every question that can come about, providing the right level of routing as needed, and importantly, collaborating with our second agent, Casey Coordinator, which is the action agent behind the scenes, taking action across many different modalities from the EMR to different system of records and ensuring that we're following not just the context of that specific home care provider but the context of the business environment overall, be it the local regulations, and ensuring that the steps we take are all in compliance. And so that's the high level of what we do. Before we go into a demo, let's look at the poll results. And so we'll have, the Anthropic team, show us that on the screen. And so here we're looking at, tremendous results, over 100 over 150 answers here, and roughly 80% of people here have or have known someone to have received home care. And why this is super important is because, as you see on that graph, folks who experience this will understand this is a problem that touches all of us, and it will continue to touch all of us. Because today, we look at home care as really this niche market for the elderly, the sick, and the disabled. But frankly, it will continue to be the place that health care pulls into more and more over time, especially as, we seek to make this logistical process of delivering home care, as seamless as possible. And the analogy I tend to draw here is that as easy as it is to order food and Amazon Prime into the home today, that's how easy we want care to home care to be for those folks who need it. Great. So we'll continue now sharing our slides. And we'll go on to the next section, which is the demo. And so I'm going to walk us through a brief demo of an example of Zingage in action. As I walk through this demo, please feel free to post questions in the chat. We'll have some time to answer that at the end. And then we'll pause at different moments to just talk about some of the underlying philosophy of how we built our agent, what we learned along the way, and how Claude has been a huge support in this journey. All right. Great. So I'm sharing here now a high level view into the Zingage ecosystem as our providers currently use our platform. The most important thing I'd point out at the onset is that Zingage operates much more like a colleague than a tool. And so what we're looking at here is not so much a tool that people are clicking around in order to achieve outcomes, but rather a transparent layer to understand what is the Zingage agent fleet doing at any given time. And the example that I want to highlight here for us is one that happens all too often in the context of home care, which is the caregiver or clinician call out. This is a situation where a last minute cancellation comes up, whether it's because of a caregiver who can't make it to the shift for a variety of reasons or, in the worst case scenarios, a no show, someone who doesn't make it to the shift at all without any notification to the provider or to the patient. And so when I click into this, what we're going to look through here is the story of what this tells and how's Zingage's action, how's Zingage's agents go to work. In this case, our caregiver, Asia, called in to report that she wouldn't be able to make it to her shift tonight with Robert. And what you'll see here is the first step being that, communication step where Riley actually receives this call from Asia. And what we're gonna do first is just listen to that call and hear the interaction. Hi, Asia. How can I help you? I had to cancel my shift due to me catching a flat tire en route to the job site, so I had to return back home on a flat tire. I'm so sorry to hear about the flat tire. That's really frustrating. Let me get the full details to make sure the team knows what happened, and that was for tonight. Yes. The one for tonight from 9PM to 5AM. Got it. So are you just canceling for tonight's shift and none of the shifts you have tomorrow? Yeah. Perfect. I'll make sure the team gets this message, and we'll go ahead and look for a replacement for your shift tonight from nine to five. Is there anything else I can help with, Asia? No. That's it. All right, I'm going to pause this here. But there are a few things I really want to point out that are special about what's going on here. So first of all, this caregiver is not using an app. They're not doing anything out of the ordinary. They're simply calling in to let the employer, in this case their home care agency, know that they won't be able to make the shift, unfortunately, because of the car issues they experienced. And what you'll notice at the onset of the call is that the agent, Riley, doesn't answer in this generic term but rather addresses that caregiver by name. Where it's possible, our agents are using deep personalization and compounding on the aspects of memory from our ability to have multiple interactions with caregivers, the patients, and all of the stakeholders. And so when they pick up the phone and call, they're not talking to a stranger. They're talking to someone who knows them and has that context. Additionally, what you'll notice is that it's asking follow-up questions necessary to really understand the situation, understand what might be necessary here for the restaffing effort. You'll see that it talks about additional shifts. Riley has the ability to look into the full schedule and understand the full understanding and context of what's at risk, what other visits might be at play here. Now, the next step really starts with taking that context and running it against the standard procedure for this particular provider's method of filling callouts. And we typically call these SOPs. Other ways of thinking about it is just a knowledge base. How does this provider typically do this on their end, if they were the one to handle that callout? And so Riley's transcript is used as the basis for the situation along with that context. And through the process of, evaluating what it needs to do next, it decides to take certain actions, and that's where Casey comes in. Casey's job now is to orchestrate the follow through of fulfilling this callout. In this case, it immediately is looking into the information in the EHR, understanding, is this a patient that can be restaffed with no questions asked, or do we need to actually check with that patient or their family? The context here for us is that you're sending a new caregiver into a home. There is a very delicate balance that's going on. We've had situations where the wrong caregiver, is sent, and people are calling the police. And so these agents have to be really great at understanding, is this something where we need to go check and get approvals, or is this someone who is a must staff, for instance? In this case, it is. It's a dementia patient, must staff an overnight, and so no questions asked, Casey goes to work identifying those next best, caregivers. Now where the agents really shine is the ability to simultaneously reach out to multiple well fit caregivers at the same time, something that an individual person can't do and just an organization with limited bandwidth simply cannot spare to do. And so as Casey is doing its outreach, it's evaluating who are the best fit caregivers, socializing this, getting them excited about the shift, in some cases even offering incentives and bonuses to pick up a more difficult shift. So think of it sort of like a dynamic pricing engine to get this shift filled. As it goes through the steps, it's reporting back again in real time what exactly is going on. And I wanna play a couple of examples here, starting with this call with the caregiver, Rebecca. Hi. This is Riley with Senior Helpers Charleston. Do you have a moment to chat? Sure. Great. I'm just reaching out because we have an urgent last minute shift that just opened up. This is overnight in Myrtle Beach, and we just wanted to see if you might be available and interested. Overnight shift. Yes. Exactly. It's tonight from 9PM to 5AM in Myrtle Beach. The client needs hands on support throughout the night. You'd be helping them with transfers using a transfer belt, help with the mobility bed, providing incontinence care if needed, and just monitoring their condition. It's an overnight hospice type of support. Is that something that you'd be able to take on tonight? I can't. I'm working tomorrow. I completely understand. That's a tough turnaround when you've got work tomorrow. So much for taking the time to talk, and I really appreciate it. Hope you have a great rest of your night. So this is a great example where the agent is not just reaching out to see if someone is available from a time perspective, but really to make sure that that individual understands what this shift entails, providing it the context of the ADLs, the care plan, while being conscious of the fact that, that caregiver, may have other things going on that are not listed in the schedule. We typically are seeing folks who have multiple jobs, and so simply looking up what their availability is for, that individual provider is not enough. And so in this case, the agent's conversation ends politely. We've captured that context, and now it moves on into, the next process. We have here the accepted shift, gives a very similar sort of perspective, and we'll play this as well. Hi. This is Riley from Senior Helpers Charleston. Do you have a moment to chat? Yeah. Great. I'm reaching out because we have an urgent last minute shift that just opened tonight in Myrtle Beach, and I wanted to see if you might be available. Yeah. Amazing. Just so you know what to expect, this is going to be an overnight shift where you'll need to remain awake the entire time. You'll be providing hands on assistance with transfers using a transfer belt, helping with bed mobility and continence care if needed, and just monitoring the client's condition overall and reporting any changes starting at 9PM and going through 5AM. Does that all work for you? Yes. 10PM. Got it. So you'd wanna start at 10PM. And would you be able to stay until 6AM? Yeah. Great. I think we can make that work, but let me just check with the caregiver you'll be relieving before I confirm. After I confirm, we'll get the schedule updated right away, and you should see the confirmation come through shortly. You can see the address and full care plan directly in the app. Thank you so much for being available on such short notice. And if there's anything else I can help you with, please let me know. No. Thank you. Alright. We'll talk soon. Take care. So we finally reach a caregiver who's accepted it, but there's a wrinkle, which is that that caregiver needs to start an hour after the original start time. And so the agent, knowing the context of this business and the context of what that particular care plan requires, pushes for the full care hours to be delivered. So rather than nine to five asking, Well, can you stay until 6 AM? And so this is super important because that is likely a 24/7 case, which means that shifting one hour can ultimately disrupt the care plan quite meaningfully, and just making sure that that caregiver who's committing to take some of the shift is willing to go an extra hour if necessary. And so the follow-up steps from here involve that documentation, which starts with updating, in this case, WellSky, making sure that the caregiver is assigned to that shift. And it also happens to be that this caregiver would be pushed into overtime for taking that shift, which would, at least in the case of this business, trigger an overtime approval request, which is not always the case. Sometimes the agency, the provider, is willing to automatically let that sort of overtime requirement go without question. But this is just a high level overview of what exactly we're doing to help autonomously power the operations and care coordination of a home care provider. And so with that, I'm going to bring it back to our slides here, and we'll have a little bit of a deeper dive into how we're using Claude to power all of this. The first one that I really want to emphasize is that semantic reasoning. There are so many interactions that we have, whether it be with distressed caregivers, concerned client family members, or patients. And words matter a lot in this business. It's not enough to just say, no, I can't make it to my shift. We need to know the context and document in such a way that folks like Medicaid can properly understand that this was a compliant method of tracking. Then we have things like consequence prediction. And so understanding what would happen if we did accept the caregiver coming in at ten versus at nine, does that mean that there's going to be an hour gap of no care? And so the agents are really trained on this model of filling in the blanks and having the situational awareness that a true care coordinator would have. Following that is the long context windows. And so you'll see things like, as the issue goes on, if we haven't filled the shift, what context do we have over time that the agent needs to keep up with so that as it continues to reach out, it is ensured that it's not forgetting something. It's not reaching back out to the same folks who have already rejected a case or perhaps have other contexts that would inform the best people to reach out to. The last few here we have are structured tool calling. And so this is super important, especially in the documentation side. We're working with many different system of records and complex system of records at that, where compliance is the name of the game here, making sure that we're documenting in the EMR correctly, that we're actually using the right tools, sending messages in the right channels, be it not including anything that is PHI to stay HIPAA compliant in something like text message and so forth. And so these structured tool calls really rely on this multi agent system having specific guardrails against calling the wrong tool at the wrong time. Next is our empathetic communication. I think that, you know, this ultimately goes back to my driving philosophy of why we're building this business, which can be summarized in a Teddy Roosevelt quote, nobody cares what you know until they know that you care. And so it's extremely important that our agents communicate that to the folks that we're interacting with, especially in such a high high intensity and vulnerable situation as home care is. Even for those caregivers who are going into oftentimes very complex situations, we want to make sure we understand them. If someone's calling out, it's not always and it's usually not their fault. And so we want to meet them where they are and make sure that we continue to build that rapport and they're trusting the agents for future interactions. And lastly is our multi agent structure. This is extremely important from the standpoint of not overloading any one agent with too many responsibilities and too many things that can ultimately overload its context windows and create a higher risk of hallucinations, higher risk of failures. And so being able to build a modular system to build multiple agents that specialize within specific domains and work together ultimately not only saves us on costs and tokens but also, of course, makes the quality of what those agents produce tremendously better. So with that, I'm excited to pass it back over to the Anthropic team and continue this conversation. And thanks, everyone, for watching our demo. Cool. Well, thank you, Victor, for your story and showing us that amazing demo. I think I speak on behalf of everyone, but what you're doing is really incredible from both the tech and the overall mission, points of view. For the next section of the webinar, I wanted to briefly cover how you all watching can build your own AI agent, whether you're building in the health care space or in different space entirely. In particular, I wanted to talk about our most powerful building block, the Cloud Agent SDK, which Zingage actually leveraged to build their AI agents, Riley and Casey. So what is the Cloud Agent SDK? In short, it's the agent harness or the agent framework that powers Cloud Code. If you're a developer, you've most likely used Cloud Code, and you're probably already familiar with how powerful it is. Cloud Code is a general purpose AI agent for software development that has access to the underlying file system, can spawn sub agents, can use skills, and much more. The Cloud Agent SDK takes all of that infrastructure of Cloud Code and enables you to build your own AI agent with the same capabilities. So out of the box, you have the same core tools, context management systems, permission frameworks, and orchestration as Cloud Code. So this is what the agent loop looks like in practice for Zingage. When a caregiver calls out of their shift, Cloud kicks off the process of finding a fillet. Claude handles everything end to end from reasoning through constraints, reaching out to replacements, adapting things as they change. First, the AI agent reasons about the constraints and evaluates the needs of the shift against complex customer specific guidelines. Then Claude starts taking actions, querying the staffing datasets, checking availability, and contacting potential candidates. The agent then weighs updates and reevaluate strategy with each new signal that comes in. Casey navigates and adapts responses, escalations, disruptions, even emergencies until the shift is ultimately filled. Victor Hunt, I wanted to invite you back on stage, but I think it'd be helpful to. How does Casey, the agent, decide who to call first? Yeah, it's a great question. So there are several ways that a home care provider thinks about who to reach out to first, depending on the demand at the moment. And so in the situation of, say, a caregiver call out, you can't quite predict when those happen. There are cycles to it where certain times of day it's more likely to have a batch. And so it's a lot to do with, first of all, who is safe and approved to send? Does this individual have the training and the licensure that they need to be able to service that particular patient? That is table stakes at a minimum. And then the next piece of that is sort of thinking about the hierarchy of needs from an acuity standpoint. If we're talking about a level four, must staff, bed bound dementia patient who needs care overnight versus someone who perhaps is like companionship, just needs a few hours of priority care, that presents a certain trade off where in a supply constrained workforce, you sort of have to think about who you're going to focus on staffing first. And so we take all of these into consideration, but some of the logical things that you'd imagine are, has this person worked with that patient before? Do they have a history with them? Is it their caregiver for Wednesday who's calling out, and we can pull in their Monday caregiver? The other aspects of that as well are geographic based, for instance. Is this someone who's ending a shift maybe a few minutes away? And so all of these pieces of information flow into the context that the agent is reasoning on and doing it at a velocity that allows it to actually get the job done quick enough to not risk ultimately losing that caregiver's ability to stop because they've moved on to something else. And so there are a lot of these sort of inputs that drive the decision making. And a big part of that, by the way, is understanding some of the strategies that the providers themselves have produced. Because these sophisticated businesses. The constraint is not their strategy. The constraint is their capacity. And so we typically onboard our customers in such a way that we understand how they go about doing this in those tough scenarios, and then the agent is able to replicate their strategy but now filling in the gaps with an unlimited capacity. Cool. Yeah. So it sounds like you guys managed to build, like, a very robust system from a lot of, like, individual different signals. I think, in general, the real world is pretty messy, and, like, people's circumstances can change very quickly. I guess, as a follow-up, how does your AI agent adapt to new signals that maybe come in, like, after, like, someone files, like, a a request for a fill in? Yeah. This is a super important point that, you know, kind of digs into the the significance of what a health care agent is doing versus almost any other context, which is that at any moment, we can receive a new signal that completely changes the care plan of that, of that issue. In the example that I shared where, the provider cancels last minute, and the agent then starts to complete outreach, it's very well possible that a caregiver who canceled last minute comes back and says, hey, actually, I've got a ride to the shift. I can come in and I can, you know, get to most of the shift. Now in order for us to, make sure that we don't burn through a number of those folks who were asking for a favor to come in and to staff this case. We have to make sure that, we create enough buffer safely so that, we can react to these new signals that come in, many of which come in simultaneously. We may have multiple people who respond yes. How quickly we're able to then get back to them depends on how much time we have to staff the case. But it's essentially doing the appropriate batching and the appropriate sort of orchestration that allows our agents in the background, Casey coordinator, to evaluate, where all of these new signals, land, and then reason on that to make a follow-up decision. And most importantly, that is sort of based on how reversible the decision is. If we're reaching out to approve someone for a shift, that is a tough thing to go back on. It's not impossible. But if someone is there on-site, they've clocked in, they're delivering care, that's sort of a set in stone situation. And telling someone, for example, the wrong information about a shift and they go in the wrong direction, that all can compound into a very nasty situation. So the agents are essentially deciding how to batch their activity on the basis of that level of reversibility and mutability, and then separate from that, sort of taking the approach proportional to the level of acuity. Yeah. Very cool. So after Casey sort of, like, does all this work and fills the shift, how does it sort of close the loop and complete the handoff? Yeah. So it sort of depends on what closing the loop actually looks like. There are certain situations where it's as simple as verifying someone is on-site, and that can be a few steps and documenting that back into the EHR appropriately, and it will auto self resolve at that point. But the way that our agents approach the problem is to understand, given the specific issue, whether it's a callout or just an emergency that comes up, starting from the end, so defining that end state and working backwards from, a plan of action, and reasoning about that planning of action iteratively. So as every step is taken, just this plan of action get us closer to that end state, the end state being this case or the visit is staffed, with documentation completed, and if that end state ever, requires, iteration itself. So along the way, we may learn, Hey, this is not just a visit that needs to be staffed, but we also need to report this to future caregivers. For example, typically, being that it's home care, it's not unusual that we receive a call from a caregiver who tells us that, Hey, the patient has passed away. I'm calling to let you know that, and the shift has effectively ended. Now at that point, it will trigger outreach to the folks who have been scheduled for the days ahead. Not just that, but the documentation steps. And so there are all these different touch points that depend on the appropriate end state, and the agent is reasoning about that. And when it gets stuck, when there are moments where it needs extra guidance, that's where it's actually bringing back in the agency and the staff themselves to help copilot and collaborate with them to understand what is the appropriate end state so that Casey can continue with its job or, if necessary, hand off entirely to the human in the loop who can then close the loop on their end. Yeah. I mean, it sounds like home care in general is pretty messy, and there's a lot of moving pieces. I feel like you can't always predict what will happen beforehand. But that that's exactly why I think Zingage product is a great use case for AI agents. Well, that concludes this section. Up next, I believe, is Q&A. So I want to invite Araba back. Hi, team. Great questions. We've had a lot of them come through. Please feel keep adding as well. To kick off, Victor, I'm curious. You're solving for a very intricate problem, as William mentioned. Can you just walk us through a little bit more on, like, how Zingage product came to be? Like, how did you know what you needed to build? Yeah, absolutely. My cofounder Daniel and I, before we wrote any line of code, we spent time directly in the seats of our customers. We worked in the back office alongside the providers who were doing this every single day, answering the calls, staying up nights and weekends to fill shifts, and, being there with them close and personal to understand how they do this today. Because, ultimately, agents are, you know, a reflection of the work that we're doing, and they unlock new capabilities as a result of what they can uniquely do different from us and from a staff. But that really started the conversation for us to understand what are the most important touch points in this workflow. How do we align what the agents do with the goals and outcomes of our customers. And so it it really starts with being up close and personal with the people who we are servicing. Awesome. I know. It seems like you guys did a lot of work to be following along and literally seeing step by step what a caregiver's day looks like. Next question we have that came in from Jack. Could you share a little bit more on when you decide to use Claude, so, like, agentic frameworks versus more deterministic workflows or, like, steps? Yeah, I love this question. It's something that when my head of AI hears this, he's going to laugh because we always talk about this. And I think the question is, you know, where does deterministic decision making fit into the context of very fluid agents? And it sort of depends on kind of your view of how these workflows come about. And basically, it's health care. There are certain guardrails and restrictions that we can clearly define up front. I sort of think about, you know, the engineering of prompt engineering as this law giving process. When you think about great law giving, you look at an artifact like the, you know, Bill of Rights, when you think about, you know, how it's structured, it's not so restrictive in the sense that it's or prescriptive rather in the sense that it is kind of giving you very clear cut rules of how you do and don't do things, you know, like Hammurabi code type stuff. Instead, it's more like, here's a framework for how you think about the reasoning behind an idea. Why might we want to document in the EHRs? Because we may have an audit from our MCOs, and if Medicaid doesn't see this appropriately, I won't get new patients allocated. And so understanding the reasoning behind this is sort of the context that gives us enough information to say, never do these things because we know that's off limits. But within the bounds of this criteria, we can allow agents to reason. And so we sort of take that layered approach where there are kind of guardrails boxed in around, the agents to prevent sort of adventuristic agents slipping out into taking actions that don't actually serve the needs and the goals of our customers. But within those constraints, it can reason about how it approaches the problem. But it's a very interesting problem because you have to imagine that these intents change all the time, the regulation changes all the time. And as the capacity increases, thanks to the agents at play here, it creates new types of considerations for what we should restrict that previously were never imagined. Awesome. I now I have a question from Zach that actually kind of flows into that. I'm guessing a lot of the people who are using your tools, like the caregivers, are have varying levels of comfortability and comfort with AI. Can you share a little bit more on how adoption has been to using these tools? Like, have you had any pushback from people interacting with them? Yeah. I suppose it's a of a give and a take where, you you have this, really great, sort of trend of AI adoption across the world. And, I I think that the stereotype of AI is that it's the calling into your airline or your bank and, you know, pressing one or just kind of going in all of these loops. And so coming out of the gate, we get a little bit of that sort of deficit at the starting line where folks are looking at all AI sort of the same. And so we have to start from a premise of trust. Where is it that people are needing us the most in communicating this is why your employer and why the provider is introducing AI is so that we can fill in those gaps where you're not getting, what you need. Before we built Riley and Casey, Zingage started with workforce engagement. We started with tools that help support the engagement layer of how providers work with their frontline workers. And by building this, we were able to deeply understand the kinds of things that are enervating to care providers, whether it be not being able to reach someone on the phone, not being able to adapt their schedules. And by understanding this, our agents had the training up front to be able to communicate, build that rapport with the caregivers and with the clients and patients and their families. And as a result, when these folks speak with the agent, they realize this is something different. This is something that is doing things in a way that they haven't seen before. A big part of that too is the memory and the personalization. So when we hear those calls answer, you know, first name, hey, William. Can you take the shift? Hey, Araba. I've got William coming over. That's the kind of personalization that gives people the trust that this isn't just the kind of, you know, off the shelf AI that most folks are experiencing. Awesome. Awesome. Next, we get a a large number of questions that come in around how you actually build the product that can be that trustworthy. Questions for both you and William. I'm gonna kind of run through some of these rapid fire. So Chuck Howe asked, underneath the agents, I'd love to understand the integrations infrastructure. So how were integrations into the various EHRs, and those are electronic health records, for each of the home care providers built. What were those 60 plus tools? How are they integrated? Yeah, absolutely. Well, our data platform is incredibly robust, and in part we've built this in partnership with some of those providers. So partners like WellSky, who have been incredible champions along the way of bringing AI to those who need it most. These are the kind of folks who have been around and built a system of records to solve these difficult problems. And so we work together with a lot of those platforms. But there are gaps. There are things that we have to build, new pipelines to understand, how to move, shifts around how to move data. And, essentially, our our goal is to make it, as seamless as possible for folks to deploy agents without having to, fill in the gaps between those different touch points. The 60 plus integrations, really vary depending on what that provider is using. An example, for instance, is that many of our customers work with Medicaid. Medicaid has a set of regulatory requirements for how you document and how you report, the actual, care that you're delivering so that you can be reimbursed and so that you are, liable fraud, essentially. And so they have specific system of records that are state mandated, and so building those custom integrations becomes table stakes for us to be able to operate with them. But all of this is to say that these integrations exist to create a bidirectional data pipeline so that our teams can readwrite into these systems, and then building the right tools so that the agents have access to communicate and to understand what to change in those systems on behalf of the providers. A question on the experience with Riley and Casey. Peter asked, how does the agent deal with interruptions? Yeah. Yeah. Interruptions is a huge it's a huge deal. It's not just, you know, the agents, frankly. I mean, we are we're on Zoom all the time. We're on, you know, phones all the time. We go through it as well. The latency is the most important thing. And so our agents are trained on the specific context of the types of things that they're reaching out and the types of individuals who they're talking with. We know, for instance, that there are certain patients who may have, a slower cadence of speaking. There are certain folks who, need to, report on a specific issue in-depth. And so knowing that context helps the agent determine how much it slows down, how much it picks up. But in general, we are sort of working from a similar playbook to some of the, latest companies, be it ElevenLabs Vappy, etcetera, to understand how do we make this feel more natural. How do we make sure that when people do need to interrupt, the agent shuts up, and it lets people talk. It gets that context. It adapts, it keeps going. But, ultimately, this is sort of a continuous progress for us, and it's great, especially when you compare it to where we were a year ago. The landscape has completely changed, it just continues to get better. Another question from Ken actually on coordination. So as you might be reaching out to multiple caregivers at the same time in parallel, we talked about this a little bit. Could you share more on how the coordinator coordinator agent makes sure that there's no overbooking? William, I might also point to you because I know that you've worked a bit on this with the agent SDK. Yeah. Victor. So overbooking yeah. I'll start. Overbooking is, you know, it's one of these, sort of privilege problems, I would call it, where the real challenge is how do you get at least one person who's a great fit, who's gonna show up in a market where you're experiencing 80% employee turnover, you're experiencing massive demand deficits or, sorry, supply deficits. And so overbooking is sort of that kind of problem that we're not as worried about. But when it does happen, it kind of goes back to William's previous question about, well, how does the agent know how to sequence the signals that it's getting? And the big thing here is really taking the time to batch things appropriately so that we are giving ourselves the best shot at doing something like filling a shift without overloading the system by getting a bunch of people excited about an opportunity and then having to reject a ton of folks as well. And so typically, in the context of home care, people understand because it's not just an agent problem. This also happens in the context of the manual process of running this business. You reach out to 10 caregivers, maybe get seven voice mails, and three people get back to you with a yes. You've already found someone else. They get it. They didn't answer the phone, and it's a matter of how well you communicate. And so being able to have that trust level and the rapport, and being empathetic to thank people for their time, for giving you, the opportunity to, maybe staff them, that's the respect that they're looking for. And, ultimately, that's what gets them coming back and understanding and excited about using working with this agency. Yeah. One thing I'll add is I think the great thing about AI agents in general is that they're very good at recovering from various error states. And I'm not saying that, like, overbooking is necessarily an error state, but the AI agent is able to reason through it and sort of decide the same way that a human would decide if, for example, you had an overbooking and you contacted multiple people and multiple people said yes. Next group of questions. A lot of them around, like, internally, how does Zingage get the most out of Claude? We saw a couple of these come in, but one from Daniel was, how does Zingage internally approach AI enablement? So do all your developers use Claude code daily? Are nontechnical teams also using Claude code? What does internal work look like at Zingage, Victor? Yeah. Everyone's using Claude. We are using Claude Code. You know, we're using Claude behind the scenes to power, you know, all of the sort of customer success and and just, even go to market. And so it's everything, that, you know, sort of scales our team. The way we look at it is that we hire incredible people. We want those people to, be force multipliers, and so we equip them with the best, resources to scale themselves. And, you simply can't do that, without, the power of something like Claude. Awesome. Awesome. Another question around trust actually came in, and I feel like it's an important point to hammer home again. So when you are, like, selling into your end customers, what do you say for for leaders and your customers who especially work, as you mentioned, in highly regulated industries or have to build, like, HIPAA compliant solutions, who feel like AI is not yet safe enough to deploy in production? Yeah. Absolutely. So we take a crawl, walk, run approach to how we deploy AI with our providers. We're very strict about the fact that our providers dictate their settings and their governance. And so while we have recommendations, we have sorts of policies and prompts that will help them deploy their agents. At the end of the day, the goal is, like I mentioned about how Claude scales our teams, is for Zingage to scale their teams. And so what are they doing correctly today, and what are the processes they have to prevent their own staff from not stepping out of bounds? And the amazing aspects of LLMs is that they can be quite responsive and consistent to respecting those policies. And so when we work with our providers, we build trust around an understanding of how they do things today and then deploying in a measured rate so that they can understand this is in fact working in a contained environment and that as we grow together with them, they can see to it that it continues to stay compliant and adherent to how they want things done. Yeah. It's interesting thinking about how you scale and, like, walk alongside your partners, but also push them to build products that they might have to, like, learn how to get the most out of. Raja had a question actually kinda similarly about how you then build that road map. So they asked, Victor, you've clearly identified a real point of friction in home care and are reducing risk in a meaningful way with the two agent model. As you look further down the road, do you see Zingage expanding into supporting caregivers more directly? So whether that's improving quality of care, reducing med pass errors, or even longer term thinking about robotics as a lever to address the human capital shortage. I'm curious how far you're thinking beyond the current surface area in terms of where you might go and where you believe the industry is likely headed to over the next few years. Yeah, absolutely. So first of all, I love this question because it's exactly why we're doing what we do, which is that our mission here at Zingage is that all health care can be in the home. And the way you do that is by solving the logistical layer of deploying health care at scale in this distributed manner. And that's what we're doing first and foremost is putting out the hair on fire problem. And through that, what we've seen is that our providers who are fully deployed with Zingage are able to redeploy their staff into new and creative ways of expanding the kind of care that they offer and improving the ability to attract caregivers, to attract partners, and overall just grow the team and the quality of care that they provide. As we continue to grow, our vision is to bring more into the care ecosystem. And we view home care not as a stepchild of health care but rather as a distribution of health care. And so where are we seeing innovation today in health care that isn't accessible in the home? That's how we look at Zingage path to broadening access and to improving outcomes. And by continuing to partner with folks like Anthropic, we can make this happen at scale very effectively. And we're super excited for what's next, be it more aspects of fulfilling operations, be it introducing new lines of care. And you mentioned robotics. I'm absolutely thrilled about what's happening in that domain and would love to talk with anybody as well who's thinking about that. Well, I feel like there are a lot of people who are gonna be excited to talk more about that. Also, a reminder to keep throwing questions in the Q&A panel. We have about five or six minutes left. I actually have a question, or if you don't mind. This is, Please. I guess, more on the technical side. But how do you guys do evals for your AI agents? Yeah. So it's very interesting because, you know, a lot of this sort of how do you know what is a good step for the agent to take, the answer to that sort of is subjective. Right? Someone might come and say, hey, it's common sense that you document in a certain way. But common sense for one group is very different from another. And so eval is a sort of collaborative process between our team, our agents, and our customers who are a part of that human in the loop process where they can actually step in and provide, essentially, ratings and commentary against what the agents are doing, which improves the memory of the agent and enriches its steps. And it's a bidirectional. It's a conversation. So these folks are looking at what the agents are doing and treating it like a colleague, correcting them. And those corrections may lead to, system wide changes in how we approach that problem overall, or it may be very specific to the individual caregiver, the individual patient. And so evals is a very collaborative process, but our job here is to absorb the context of our customers so that as the agent encounters new scenarios it hasn't before, it is able to react to them similar to or even better than the providers themselves would have. Yeah. Makes sense. I guess so I I've been seeing this a lot in the Q&A questions, but some people are asking, like, what differentiates you guys from potential competitors, and how has perhaps working with Anthropic helped you separate yourselves from the rest of your, I guess, other companies that are also doing something similar? Yeah, absolutely. Well, when we look at a lot of what's out there, there are a lot of point solutions that we've seen touch different parts of the workflow. What Zingage really focuses on is building a platform that is extensible enough to meet providers exactly where they are, and that's why we've built our modularized agent our modularized agent SKUs where you can deploy these agents on specific problems as you need, and they continue to adapt as opposed to solving for a single type of a problem such as, you know, filling a call out. That is a very common and painful problem in this space, but it's not the only problem. And on top of that, when you look at the steps that it takes to successfully complete it, that is very varied from provider to provider. And so when we look at what's out there, we are excited about how folks are thinking about just solving for the space because the space needs a lot of attention. But what we care the most about is, is this going to adapt and grow with the dynamism of the space? Because we're working with folks and people's lives here at the end of the day. And so we need to be able to react to a very volatile health system. So that's the biggest difference, is we're a platform, and I see a lot of point solutions, but not very much on the platform side. Yeah. Yeah. So that goes back a lot also to what you were saying earlier about building trust and building a strong brand around trust. I think we have time for one or two more questions. I wanna pick back up also the thread, William, that you just introduced around evals. I'm curious for you, William. You work with a ton of different partners. What are the markers of good evals? Like, what are best practices, especially as a lot of the folks listening in might be starting to build out their even offsets for the first time? Yeah. That's a great question. So we recently published a blog post about how we at Anthropic think about evals. And I think the sort of, like, best way to segment evals is you kind of think of it sort of like software testing in general. You have your, like, unit tests, which are evals that sort of evaluate tool calls. Like, are the inputs correct? Are the right tools being called at the right moments? And then you have your more, like, integration and end to end tests, which are evals that just sort of, like, test the overall behavior of the AI agent. So is the final output correct? Are, like, the intermediate outputs correct? Are, like, the correct sequence of tools being called? And I think over time, you can take, like, real world use cases for your AI agent and gradually turn more and more of those into individual evals. And you can take failure cases or, like, things that you want your agent to be eventually good at. And then if you build evals around them now, you can eventually, like, get to a point as you build up your system where you can, like, verify that what whatever you're building is actually, like, what you want your AI agent to be able to do. So I think it's a very hard process. It's very iterative, but I think it's very important when you're building AI agents, at the frontier, because a lot of the times, like, capabilities evolve. And I think having robust evals allows you to, like, move much faster than companies that don't have robust evals. Awesome. I feel like back to being a differentiator, especially, Victor, with what you were mentioning. It seems like your eval sets help you guys test and iterate and build more intentionally and effectively. I know we're coming up right back on time, and I see that there are a lot more questions. But thank you so much for your insights, for your candidness, also just for sharing such a personal story, Victor. This wraps up today's webinar with Zingage. To keep the conversation going, we're actually gonna be sending a few resources through email after the webinar. So there'll be a recording of what we covered, answers to the questions that we did not have time to get to, and then a few key resources to help you get started. So we'll be sure to tag the eval docs that William mentioned and then our agents ace SDK docs as well. With that, enjoy the rest of your day. Thank you all.