We sat down with Hyro’s CEO, Israel Krush, to discuss his journey and the company's mission to simplify interactions across all digital channels, so that companies and their customers can communicate easily and achieve more, together.
Hanna Edgren: Could you tell us about your background and how you got into the world of conversational AI?
Israel Krush: I started my career in the army, specifically at the Elite Intelligence Unit 8200, where a lot of other Israeli entrepreneurs come from. There, most of these entrepreneurs go into cyber because Israel is notorious in that field. When you think about intelligence, there are natural language understanding aspects to it. At its core, it is about dealing with big data; how to extract and filter it for operational needs.
My background is in computer science and statistics, so I like to say machine learning before people called it machine learning. I was a software engineer at Intel and later at a bunch of startups from cyber security to ad tech. I have spent the past 20 years of my life in the Israeli tech scene. In my last role, I was the Head of Product for a computer vision company that was recently acquired by Walmart, so I would say the thread connecting my experience from the army to today is dealing with human computer interaction (HCI). Whether it is language or vision, I have explored how we as humans communicate with technology.
In 2017, I moved to the US to get my MBA from Cornell University, specifically developing the new campus in New York City called Cornell Tech, which as the name suggests, focuses on technology. That is where I met my co-founder, Rom Cohen, who was doing his Masters in Computer Science at Cornell Tech. We met and were exposed to the voice space back then. We were talking about the entrance of Alexa, Google Home, and all these devices that were still not big in Israel. We were excited, so we purchased a few devices, tried them, and were disappointed.
Hanna Edgren: A fascinating background and a great segue to my next question. What led you to founding Hyro, and what problems are you looking to solve?
Israel Krush: Trying these devices led us to a few big hypotheses that stand at the base of Hyro and what led to us creating this company. One, which is probably more obvious today, but in 2017 was less obvious, is that natural language interfaces are going to be the dominant interface when it comes to human-computer interaction. We are going to communicate and interact with technology via natural language interfaces whether these are voice-based, think Alexa, all the voice assistants in call centers, or even SMS-based correspondence. At least in the tech industry, this is not something that people will be shocked by, mostly due to recent advancements like ChatGPT and other large language models.
The second thing that led to Hyro, was that these systems are hard to deploy and maintain, whether this is a chatbot on a website or a voice assistant in a call center. It is especially difficult if you are a large enterprise with a lot of data from various data sources that are often siloed. I am not going to dive too much into the tech, but I will give you an example. Let’s say we want to help patients find physicians based on various attributes. When you design an AI assistant, there’s a very important concept called intents. You can think about this as a micro goal that the user might have, so in this case the intent would be to find a physician based on their specialty, another intent would be finding a physician based on their insurance plan, and the third intent would be the combination of these two things. You need to predefine or think about all these intents. The more data you have, the harder this becomes as the number of intents grows exponentially. For the final position use case, you sort through dozens, if not hundreds, of different intents. Then, you need to create matching conversational flows for these intents. Given that they identified your specific intent, we ask, what are the questions that I need to continue to ask you? Finally, how do you classify the relevant intent based on natural language utterance, whether typed or spoken. This is done via machine learning which requires you to gather lots of data, train models, and deploy these. In doing this, you learn that your users are asking completely different questions than what you have intended, so you fine-tune and retrain the model. This reliance on machine learning and modeling each use case with this intense architecture is hard both to deploy and maintain, and most organizations failed in doing so.
What we have done is take an entirely different approach by mimicking what really happens in a call center environment with humans. Again, for the final physician use case, probably the main knowledge would come from the physician directory, which are sometimes available on the website. We tap into the organization’s existing data sources whether that be databases, APIs, CRMs, etc., and we scan them to construct the knowledge graph. This knowledge graph is a type of data structure that shows main entities, their attributes, and relations between those. In the physician use case, the main entity would be a physician, and some of the attributes would be the specialty, the insurance plans they accept, location, etc. We can take this user utterance and break it down to understand the different words and their role in the sentence based on composition and grammar. This is referred to as semantic meaning. So, understanding the semantic meaning and being able to traverse the knowledge graph we build allow Hyro to retrieve the relevant information. That is what makes us unique in this space. We are focused on shortening the deployment and making maintenance almost 0-effort. We can arrive at a first meeting with a demo ready that is built from this customer’s website taking anything publicly available. This opens a lot of doors for us and helps us immensely in the sales process.
Hanna Edgren: That is great. Can you share a case in which Hyro’s technology has made a big impact?
Israel Krush: A few years back, we were one of the first in the world to recognize that there was this new thing a lot of patients were asking about, being COVID-19. We were able to give health networks, HMOs, and hospitals a chatbot that was able to answer frequently asked questions about COVID and do a self-assessment based on trusted resources. This was a time when health networks were so busy treating patients that they didn’t have any information about COVID on the website, and even if they did, updates were coming out so frequently that they often displayed old guidelines. This is the best example of deployment and maintenance that would be very time consuming for these organizations. We were able to tap into resources and digest this information as it was updated. We put chatbots in almost two dozen health networks in a matter of days, the fastest being 48 hours. There was a huge spike in call center traffic, and with this assistant we were able to reduce 30 to 40% of the calls to call centers.
We haven’t yet talked about the analytics component of the solution, but we aggregate all these conversations into insights that help with two main things. One of these is giving insights to customers. So, 70% of your patients are looking for cardiologists in Brooklyn, but there is 0 availability in the next 90 days, for example. The other thing that we do with this layer is identify knowledge gaps. A few years ago, we identified a spike in our misunderstanding because of 1 noun, which you can guess is ‘booster’. Suddenly patients started asking about booster shots, and we were able to find available information and answer these inquiries. This case was one of our most impactful, helping tens of thousands of patients and health systems that were very overloaded. Because of this work, we won the Microsoft Partner of the Year Award. That recognition was exciting and truly a mark that our work was resonating.
Hanna Edgren: In the early days, how did you think about product market fit and how did you know when you had found it with Hyro in healthcare?
Israel Krush: Pre-COVID, we saw this industry as one that is ready for digital transformation with evolving patient expectations. COVID sped up this digital transformation by 10 years, so with the case I described previously, we really felt like we had found our market. As we dove deeper into healthcare, we realized a lot of inefficiencies that happen both on the provider and payer side. Even with simple administrative tasks such as scheduling an appointment, patients were frustrated. This became clear by talking with customers and end users. It really is a win-win situation. It is a win for the patient because they can text or call 24/7, and for the organization it is about increasing operational efficiency, reducing burnout from staff that comes from doing these repetitive tasks, and at the end of the day cost savings. For us, we are doing impactful work and getting paid for it, so it is really a win-win-win.
Hanna Edgren: The healthcare industry is complex, and privacy and security is a big part of this complexity. How do you approach this at Hyro?
Israel Krush: I think that this is one of the most important and underrated challenges, at least in our category of natural language understanding, especially with all the innovation in the space. Letting an AI decide what to say or write to patients is problematic, and they will have hallucinations. “Hallucinations” basically mean making things up, and this contributes to the alignment problem. ChatGPT is called ChatGPT because one of the goals is to be able to chat with you, so it will prefer chatting with you and making things up over saying “I don’t know”. Everything that I am sending to OpenAI is personal identifiable information (PII) that can be sent to the cloud. There is existing compliance, like HIPAA, which is one layer, and the other is future regulation. People are talking about the need to regulate AI because it is a huge power, not only in healthcare.
We are working to preserve everything related to HIPAA, high trust, and GDPR. Besides that, we are working on 2 main mechanisms that will help with future regulation and compliance, which are around explainability. How can we explain to our customers and end users why the AI gave the answer it did? That is very hard to answer with machine learning because it has become a black box, in which people don’t understand what is going on in between the input and output. The second mechanism is around control, so how are we able to give control to our customers? We want them to be able to say this dataset is okay for this patient but not okay for that patient.
Hanna Edgren: Could you talk about your vision for the future of AI?
Israel Krush: I think the pace of innovation around large language models is outstanding. It is something we haven’t seen in the past. The regulation does not keep up, and potentially there are going to be some bad players, and of course, good players. Where I see conversational AI developing is through industry-specific use cases that either require proprietary data that makes your model better or some techniques that you are using like knowledge graphs that give unique attributes. At Hyro, we are heavily focused on regulated industries like healthcare and insurance which will require a different type of solution. My hypothesis for the future of conversational AI is that it will be very much verticalized. You will have some infrastructure software such as these large language models from Google, Meta, or Microsoft. But most companies that will be built will be verticalized and focus on specific aspects of this vertical through unique data or methods they use.
Hanna Edgren: That is a very interesting hypothesis. Now finally, what excites you most about leading Hyro, and what are you looking forward to in the journey ahead?
Israel Krush: What excites me most is that every day I have a new challenge. I would say we are lucky enough to be in the right place at the right time. We are now rolling out Spot, which is a new product we have that is GPT powered. The product can take all the information on health network websites and digest it in a way that is explainable and can function like a plug-and-play. This level of interaction in terms of the natural language experience is incredible to the point that you think you are speaking with a highly skilled human. We are launching this in congruence with the announcement of our $20M Series B with investors Macquarie Capital, Liberty Mutual Strategic Ventures, Black Opal Ventures, and more.
Comments