Štěpán Kopřiva (Blindspot Solutions): When a person reports being quarantined in the morning, we can reschedule shifts in seconds thanks to artificial intelligence

• How do Czech companies approach AI, and what burns them the most?
• Conversely, what prevents its greater spread?
• Can algorithms be used in processes where frequent random changes occur?
• Why involve AI solutions even in an area such as packaging management?
• And how does an AI-driven discount mechanism work in retail?

Ivana Karhanová: How do Czech companies approach artificial intelligence? What solutions do they demand? And what are they most interested in? What is preventing a greater spread of artificial intelligence? That's what I'm going to talk about today with Štěpán Kopřiva, CEO and co-founder of Blindspot Solutions, part of the Adastra Group.

Štěpán Kopřiva: Hi.

Ivana Karhanová: Blindspot Solutions specializes in solutions that use artificial intelligence. What do Czech companies demand the most?

Štěpán Kopřiva: I think the spread of artificial intelligence is, let's say, in the middle now. It's not like what we see in Western markets. First, there is much demand for solutions aimed at logistics and production, that is, their management and reasoning. Then some solutions are used to save human work - things related to customer support, automatic messaging, automatic reading and processing of documents, or solutions leading to the prediction of future events.

Ivana Karhanová: You mentioned that artificial intelligence is more widespread in the West. The European Commission even released the Digital Economy and Society Index 2021 in November, and the Czech Republic is ranked a nice 18th out of the European 27. So why aren't Czech companies embracing it so much yet?

Štěpán Kopřiva: I think it's because the Czech market is a bit behind the Western market and hasn't had this need in history. It's also influenced by covid. However, nowadays it is proving to be a problem to find employees. This means that the motivation in AI is no longer there. We just want to save money, innovate or get something new so that companies can function at all. I think that is one of the reasons. Maybe the other driver is that many companies in the Czech Republic have their mother somewhere else, and the adoption of the trend is slowing down for that reason as well.

Ivana Karhanová: What is natural, artificial intelligence in the eyes of Blindspot Solutions? And what are the solutions that use machine learning?

Štěpán Kopřiva: From our point of view, machine learning belongs to artificial intelligence. The software can learn and generalize something is an essential word for us. It means that if the software can solve an example that it hasn't seen before, it's artificial intelligence, in our opinion. But, of course, other things go along with that, like scalability and, to some extent, the impact that it provides.

Ivana Karhanová: Can you give an example of that?

Štěpán Kopřiva: We can elaborate on this with the example of computer vision. For instance, we have a solution called Macula. When we train an algorithm and want the program to recognize cats, for example, we show our model or program a training set of cats. It learns to recognize pictures with cats and photographs that don't have cats. We teach the program to do this, and later when we show it a picture of a cat that was not in the test set, and it recognizes it correctly, we say that it can generalize. So he's giving a result for something he's never seen before.

Ivana Karhanová: You used Macula to identify drapes during the lockdown.

Štěpán Kopřiva: Yes. We have used Macula to detect drapes in retail, where the solution both detects whether or not a person is wearing a drape and counts customers in the store. We trained it on faces. So when the solution detects in someone new whether or not they have a drape and can even identify a drape from a scarf, for example, that's precisely the generalization we're talking about.

Ivana Karhanová: So that's already artificial intelligence?

Štěpán Kopřiva: Yes. It has to be said that it is a specific artificial intelligence. It's not something that is called general AI. That is something that companies like Google are working on. Nobody knows if we'll ever achieve it.

Ivana Karhanová: Is that the AI we should potentially be afraid of?

Štěpán Kopřiva: Yes. But I'm not afraid. On the contrary, I'm looking forward to it. If we ever get to it, for me, it will be a big step in the history of mankind. The question is whether we will even make it.

Ivana Karhanová: Let's look at how Blindspot approaches the tasks that the market demands. You mentioned examples of use cases in logistics and optimization. For example, you've done work for Skoda where your algorithms tell you how to load a container.

Štěpán Kopřiva: Yes, that's right. We have the Optimus solution. Specifically, in the project for Škoda Auto, our algorithms tell how to load the pallets into the container to make the most of the space. The problem is that the pallets are of different widths, heights, and weights. Therefore, it is not humanly possible to plan how to stack the pallets and put them into the container in real-time. This is one of the challenges we are addressing.

Ivana Karhanová: We should say that without your algorithm, they were able to fill the containers, but with your solution, they can save about three more cubic meters of space.

Štěpán Kopřiva: Yes. The algorithm saved several cubic meters. In terms of the total volume of containers and terms of money, these are very interesting numbers for one year. So the impact is really visible, not only in business terms but also in terms of responsible and ecological behavior, because it saves a lot of CO2 emissions at the same time.

Ivana Karhanová: When we talked about the solution before the podcast, it was also mentioned that it also helps them in that when a new person comes in, they don't have to learn stacks of papers on how to stack the container. How long does it take to deploy a solution like that?

Štěpán Kopřiva: It depends if it is customized or if the company uses one of our products. We currently have the Optimus platform, which can solve planning and scheduling tasks involving large resource pools. For example, it's route planning in internal or external logistics for trucks. The solution makes sense even for truck counts in the order of higher units (8, 9, 10, 15). The solution can tell which truck should go when and where to use the optimal number of trucks. This means, for example, that trucks can take packaging materials to the production line and finish products straight from the production line. All this can happen within a perimeter of five kilometers, for example. If I translate this into a situation before the software was deployed, let's say that 15 trucks were running before. After deploying the software, we have experienced from several clients that the number of trucks, drivers, and costs can be reduced in the low tens of percentages, maybe 10 to 20%, which is a significant saving over a few years.

Ivana Karhanová: What all can be planned with the Optimus solution?

Štěpán Kopřiva: We plan resources such as cars or people. For example, we can plan when each person should come on shift, especially in more complex operations, factories, or even offices where not everyone can do everything. It's a complex task, especially when, for example, there are many agency workers involved. For example, for our client in Germany, we handled the scheduling of their inspectors in China. Their inspectors go to factories on the East Coast, and we plan through software who should go where, when and who should go where. We can plan shifts for large numbers of people in real-time and reschedule random shifts.

Ivana Karhanová: Like when a person tests positive for covid.

Štěpán Kopřiva: Exactly. Or when he had a Christmas party :-) The software can find a new solution in tens of seconds. We also plan things from the physical world that doesn't seem so important at first glance. For example, we're able to determine the size of the packaging. It's one of the projects we tackled for a prominent client in Germany in the cosmetics industry. He has an e-shop that sends different shipments. The problem is what to pack the shipment in if he has five different types of boxes. In addition, in delivery, they pay the price according to the size of the box, and the demand for the goods can change over time. The solution we have developed can reschedule box sizes to match product demand, for example, monthly or quarterly.

Ivana Karhanová: Do I need artificial intelligence for this? I imagine that I take the products we have sold, for example, in a quarter, somehow calculate it and order the right boxes.

Štěpán Kopřiva: Yes. The problem is that the boxes can be customized. The second thing is that there are so many orders and so much data that it takes a long time to calculate manually. There are too many combinations in that volume. And that's a feature of most of our projects. We may improve the result by units or small tens of percentages, but it can make several million euros per year in total volume.

Ivana Karhanová: You said that you use planning not only in the physical world.

Štěpán Kopřiva: We can also use it in the virtual world. For example, suppose someone uses multiple cloud solutions or providers or uses different computing machines and has different computing tasks. In that case, we can use the software to schedule which machine should run which tasks.

Ivana Karhanová: Does that mean using virtual servers and things like that?

Štěpán Kopřiva: Exactly. I'll give you a simple example. A customer has two kinds of clouds. One is internal computers; one is a cloud from a larger provider. The internal computers cost less, but it's not scalable. He has maybe ten internal computers, and he needs to do some computational tasks - typically scientific calculations or predictions. Those jobs can be sent either to internal machines or to the cloud. It costs a lot more in the cloud. And when there are relatively many jobs, and we know how long it takes to count them, and when we need the result, the software can tell us where it's worthwhile, whether it's in-house or on the cloud. So it optimizes costs.

Ivana Karhanová: By using artificial intelligence, can I solve cost management?

Štěpán Kopřiva: Exactly, but under certain conditions. The plan guarantees that each task is calculated exactly by the time the result is needed, while at the same time, it all costs the least amount of money.

Ivana Karhanová: I'm able to eliminate bots in cloud cost management when I might have machines left running somewhere that I don't need.

Štěpán Kopřiva: We don't completely deal with this part. We're not dealing with the actual detection of the running machines on the cloud but where the computation should occur.

Ivana Karhanová: No, we don't. I thought the algorithm would say, "Now send it to this server and then shut it down."

Štěpán Kopřiva: Yes, it does. It's always working with big data. So we're using cloud automation, machine dialing, infrastructure service, etc. Our solution tells where to drop the job.

Ivana Karhanová: Another solution that uses AI is customer support. Companies are addressing that there are no people on the market to put in all their call centers. How does artificial intelligence help here?

Štěpán Kopřiva: Artificial intelligence helps by automatically replying to text messages. We have a solution called Aris that can classify and sort incoming emails. I'll explain it using the example of an e-shop that receives many emails like "Hello, I didn't get my package. And so on... Aris can classify the message and prepare a reply, and the customer support representative spends much less time processing one email. We go from tens of seconds or a minute to somewhere in the units of seconds - to seven or eight seconds. The customer support tells you which reply is correct and has it sent. And that's not to mention that some messages may be answered automatically.

The key is that the easy to reply to messages are tens of percent. For example, we can handle 40 to 50% of messages almost automatically, and the savings are huge because of that. For example, it's the shopping season before Christmas, and everyone in e-commerce is figuring out how to have enough people in the customer center. We can talk about hundreds of people for one customer. And when suddenly half the inquiries are resolved in almost zero time, you only need half the people.

Ivana Karhanová: In the case of these text message responses, are we talking about AI or machine learning?

Štěpán Kopřiva: It's machine learning, specifically NLP - natural language processing. NLP has gone well ahead because of the proliferation of deep learning and transfer learning, and that's what we're building our products on.

Štěpán Kopřiva: How long would it take you to deploy such a platform to a client?

Štěpán Kopřiva: When a client comes to us and gives us the historical data, we are learning the system. It is units of days. During that time, we can dial up an instance on the cloud and prepare it for the client to see how the system works on their data. The integration into real operations then depends on the client's internal systems. Even with integration, it can take a few weeks. So it's not a matter of years or long months or anything like that.

Ivana Karhanová: Blindspot also predicts demand. It's clear that when it rains, umbrellas will sell. When it's a holiday, and it's nice, fewer people will be in the shops. That's probably very trivial. What are you able to predict?

Štěpán Kopřiva: We usually predict some variables that depend on other variables. In shopping, the day of the week, the hour, the weather, the time of the year, or even if there is an event in town, we can incorporate almost any data that is available in machine form. When I look at what customers want us to do, it's typically demand for goods. How many goods are sold affects the need for staff in warehouses, the delivery of goods, the need for cars, etc. That's where demand forecasting and optimization, planning, and decision-making come together. We can predict how much goods will be sold and, based on that, optimize the need for people, cars, or the distribution of working shifts.

Ivana Karhanová: Theoretically, it can also work by minimizing wasteful chains and similar tasks.

Štěpán Kopřiva: Exactly. We also deal with tasks like pricing goods that have a shelf life, typically food in a warehouse. The chain would like to know how to sell the food before it expires. The mechanism to do this and speed up sales is called discounting. The chain asks how much to discount the item so that it ideally sells it all. Anything they don't sell goes straight into the negative.

Ivana Karhanová: This then influences all my business processes further down the line - warehouse planning, shift planning, logistics planning, etc.

Štěpán Kopřiva: Yes, it can be. We learn models on historical data, and the inputs are real data in real-time. It always depends on how many items are affected by such a discount event. All of these businesses are set up for discount events. They routinely have promotions, so their systems are set up for it. This is just another input into the discounting process.

Ivana Karhanová: It does not just demand that needs to be predicted. You also predict churn. I suppose especially in services. What all goes into the algorithm?

Štěpán Kopřiva: We do it, for example, with manufacturers and service providers, often digital services. Thanks to this, the operator or service provider can react in some way, such as giving a discount or contacting the customer and asking what is wrong. Data comes into it that the operator, the provider, and/or we can measure. To give an example of a digital service or app, we can measure things like how often a person logs into the app, how much time they spend there, how much time they spend in each section, how many clicks they make, and how long the messages they send are. Well, and when we're able to measure things like that, we can identify a behavior change.

Ivana Karhanová: For example, he starts browsing certain sections of the site where no one has seen him before.

Štěpán Kopřiva: For example. Or suddenly, he doesn't go to the app at all, or he spends a smaller part of the time there, and so on. We call it anomaly detection. We have a long experience with that. We detect anomalies in cyberspace and in the physical world, for example, anomalies in production. For example, it's a chemical process. Something was made a little differently than it should have been, so it's suspicious. We're able to detect it automatically. Or we can detect anomalies in other processes where data is collected. For example, in our company, we collect many data on how much time our people spend on projects, how much code is written, etc. Even on things like that, we can see if there's something different going on. It doesn't always mean something is wrong, but it does alert us. So in companies where data is collected on almost any process (HR, back-office processes towards invoice processing, etc.), we can plug the data into the software and look for areas where something strange happens. This can be of value to a manager or anyone who owns the process and wants to monitor it.

Ivana Karhanová: We mentioned churn prediction. What else can be detected this way? Anomalies are an awfully broad term.

Štěpán Kopřiva: It's a broad term. For example, when someone is making moldings out of plastic in manufacturing, we can detect anomalies by visual inspection. With a camera, we can see something strange about the molding. That's one example. The second example I will give is from the digital world, where we can detect, for example, anomalies in the re-endorsement process. Or how someone is reporting their costs. For example, an employee shows a certain cost over a long period. For example, he shows ten thousand crowns for a bigger working dinner twice a month, and his manager has to approve it. But he may have some limits that the manager no longer approves, and the system suddenly shows that the employee has started uncharacteristically reporting five hundred crowns almost every day. And that's a behavior change. The system will allow the manager to keep track without tracking individual data. That's the problem today - there is many data, and managers cannot track it all.

Ivana Karhanová: You don't even want to control or suspect people.

Štěpán Kopřiva: Exactly. And it's not just about that. The system helps by picking subsets out of many data. It picks a few cases that it says are weird. And the whole thing runs automatically somewhere in the background. A manager might get a summary once a day in an email or an app. He gets a description of the anomaly, and that way, he only deals with the important cases. It can also be used for processes that are not completely critical. If we use the Pareto rule, 20% of the things create 80% of the value, while the majority is not that important for the manager, and the system can do those things for him.

Ivana Karhanová: In these tasks, I don't have to set any boundaries for the algorithm as to what it has to watch and what it doesn't have to watch. The algorithm should be able to do that itself. Otherwise, there is always a way around the system.

Štěpán Kopřiva: That's exactly right. If we are thinking about bypassing the system, we can discuss some fraud detection. Still, it can also be in production when the characteristics of the machine change somehow, for example, the temperature. Our system doesn't need to set boundaries, and it detects anomalies.

Ivana Karhanová: During today's podcast, we've mentioned many examples where algorithms in generalization can help with basically anything. But some tasks can be complex as a result. So what do you recommend companies start with?

Štěpán Kopřiva: We do an AI assessment, during which we advise companies on how to bring AI into the company. That's where I recommend starting. I also talk about companies with tens of thousands of employees and are foreign. We advise on strategy, and our assessment includes assessing the quality of the data collected. Then, we proceed based on the benefits of implementing a particular system and the complexity. Because when a system is implemented, there is a benefit, but we often find that the problem is more on the technical side - how to make the system work.

It is also about the human side - getting people to use the system. This means that we have to show our solution to people so that they can trust it. Because our solutions are often found in ways that are different from what people are used to doing. We also usually deliver the solution in the form of a software application, which, in addition to functionality, must be user-friendly so that people find it easy to work with.

Ivana Karhanová: When you agree on a POC (proof-of-concept) with a company, how long does it take to implement the solution?

Štěpán Kopřiva: It depends on how dynamic the customer is. If an assessment is made, we can start the POC within a month. The proof of concept can then take another six weeks or a month. We're used to working dynamically, so we're not discussing years of creating theoretical strategies.

Ivana Karhanová: So we're talking about months. And what about the payback?

Štěpán Kopřiva: The return depends on the domain, ranging from units to months. We have had cases where the return has been 3-4 months to 1-2 years. When the return is longer than two years, it is problematic for a company to invest in it nowadays. Then there are larger projects, which have to be approached a little differently.

Ivana Karhanová: Thanks for talking to us today. That was Stepan Kopriva, CEO, and co-founder of Blindspot Solutions. Thanks for coming by the studio.

Štěpán Kopřiva: I thank you so much for having me. It was really interesting.