Introduction

So the main topic is what do you need, in fact, for a successful AI project?

Understanding the Benefits of AI

And so if you go to the benefits of AI, I think it's important maybe to stress two main benefits of AI.

The Evolution of AI

In fact, AI is nothing new.

It exists already for many years.

AI vs. Traditional Models

But if you compare artificial intelligence to econometrics or to normal mathematical algorithms you use, the difference is that artificial intelligence algorithms help you to find

invisible connections and big amounts of data.

So you use neural networks and when you put data inside, they look for invisible connections between this data, something that you can't do with mathematical models.

And this is like the main difference and it uncovers at the same time some blind spots.

So if you have like massive data,

with a neural network you can find these invisible connections.

Case Study: Invisible Connections in Data

In my further slides I will show you an example of a project where we've done that and it made a big difference to mathematical modeling.

Another big advantage is that neural networks train all the time, so when you have an artificial intelligence AI model,

And you have a process going on, so we can still retrain it and make it better with real data.

So we are constantly improving.

So we have a model, then you have your real data.

For example, if you do sales forecasting and you have your real data on sales, you are all the time constantly retraining your models.

And that's why they become better and better and even more precise.

Case Studies in AI Applications

Predictive Maintenance for Windmills

To show you some examples, this is, for example, a case we've done for windmills.

In Denmark, and in this case, like anybody of you knows windmills, all of them have SCADA.

SCADA is like a system which analyzes a lot of data out of a windmill.

So it analyzes about 140 different parameters.

And normally, when there is a difficulty, for example, if an oil state...

becomes too low and like a temperature goes high for example in the connection between like this rotor blades and the gondola then the windmill goes in a state of maintenance so it doesn't function because it's too dangerous but with SCADA you see this already when it happens so you see when this

dangerous state already happened because there was like a severe wind and there was like high the the temperature went up and the windmill went into a maintenance state but if you use so in this case we we used high speed sgtm neural network and we in fact we were we had like data of good states of the of the windmill so the time when everything was good with the windmill we trained the model

And with this model, we could already receive a warning that the windmill will go into maintenance state already 48 hours before.

So we got an information 48 hours before.

the oil state in a certain place goes down.

So it means at a certain point, the windmill will go into maintenance state.

But if you repair this, so if you, for example, fill in the oil, then nothing will happen.

And this means that if you use...

neural networks, and if you use the eye, you can predict these maintenance states and avoid them.

1So it allows you to have the functioning time of a windmill much longer than before.

And this is the main difference, that you see these invisible connections, because what you won't see with a normal mathematical model.

Weight Measurement of Pigs with AI

Another case, which I personally find very funny, but it's in fact the most complex AI project we've ever done.

So we had a task to measure the weight of pigs using AI.

So normally if you have like a pig farm, all pigs, okay, it's like unfortunate, but at some point they go to a slaughter.

And when they go to a slaughter house, they have to have a certain range of weight.

And because otherwise the slaughter needs to adjust knives and it's like,

It's in fact a time waste, and that's why the pig farm pays some penalties if that happens.

That's why all of them weigh pigs, because they need to have a certain range of weight.

They do this manually, so normally it takes two people and half an hour per pig.

But if you, so the idea here was to develop like a device, it's a camera device, and with this device you make, so you see it's like prototype in the hand of a man there.

So we make photos of a pig, and using AI algorithm, we detect, so we like in fact calculate the weight of a pig based on these photos.

And so it's like quite complex scheme behind.

But at the end, we have an accuracy of about 95% on this weight monitoring.

And in this case, we in fact save.

So it's like we make the whole process much more better for pigs because they do not notice that something will happen to them because we simply make photos and we are not trying to put them on a scale.

And at the same time, you can save time because you do this much more quicker than if you use a normal method.

And this is also something where you can use AI.

So maybe nobody would think, OK, this is a case or a business case.

But also in such cases, AI is helpful.

This is like more industry-like case.

Predictive Maintenance in Industrial Settings

So we've done that for Sensoro.

It's like a German startup.

And in this case, it's a predictive maintenance, in fact, for rolling bearings.

So if you have like a rolling bearing in a big industrial machine inside, and if something happened to this rolling bearing, so if it's out of use, so we need some time to replace it.

And it's always like, it puts the whole machine on hold.

And in our case, we use, so we measure ultrasound noises.

And we have, there is a device which measures this ultrasound noise of rolling bearing.

And we put then the signal from analog to digital and use LI algorithms to detect if there is any brokerage of this rolling bearing or something not really normal, how it should function.

And with this detection, you can detect, OK, this rolling bearing will maybe go broken in the next time.

And you can replace it out of, like, in fact, on needs and not on time, how it normally functions.

AI-Driven Sales Forecasting

Like, a bigger case I brought in which I find, like, very good for different industries is AI sales for recasting.

So we've done that project for a fresh meat chain.

And their main challenge was...

so imagine if you have like a fresh meat shop and you sell fresh meat and if you do your ordering or like your pre-orders not correctly it means you've pre-ordered like too much of pork meat and at the end you haven't sold that in the next days so you've generated like waste because you can't keep it like too long

And so you have this, for example, also at bakery shops or at other shops which sell products which got spoiled quickly.

And here the aim was to use AI to predict sales and with these predictions to make pre-orders better and more precise.

So our main goal was to reduce waste of this fresh meat chain.

But at the end, we also improved their product matrix.

So when you see here, so when you have manual orders, you have a lot of different factors.

For example, there is a human factor that if you have a person who makes these orders which is new to the shop and has no experience, like normally,

possibility that he or she makes a mistake is quite high because she has no experience.

And if there is an AI model behind, it will give her a suggestion.

For this day, for example, for Tuesday in this month, you should pre-order so much of this and this meat.

And so it gives you a much better idea what to pre-order.

And at the same time, it's also quicker, so you don't need so much time for pre-ordering.

But also interesting here, what can you do better, in fact, with AI in comparison to how they've done that before?

So as you see, AI gives you a possibility to adjust future sales.

And if you use like normal econometric models, which they were using before, you always see like the past.

So you see the sales

like out of like past three months for they were like comparing always three months periods and at the end you can't improve this three months so if you've sold like if you had like bad orders so you can't improve them anymore and if you say hi it's always like future oriented so it gives you suggestions for the future and you can like you can for sure also correct on your own because the the main

challenge we had in this project that you can predict everything which is regular, but if you have some irregular sales, for example, somebody is

celebrating a wedding or there is like a big big birthday party planned and you know that somebody came to your shop and pre-ordered for saturday like 40 kilograms of pork meat it's something that you can't i can't know because it's like an it's not a regular regular thing but this you can add to the suggestion of ai so you can still validate the suggestion and add this this individual things

Methodology and Results

These are some facts about this project.

So if somebody is more deep into AI, we used two approaches here.

So time series forecasting for normal, regular things, and prediction with the regression models.

If you had some...

some periods of discounts, so if there were some advertising times or some discounts, we use for that prediction with the regression model to correct our time series for recasting.

And also, at the end, we had five test shops.

And at the end, we managed to reduce waste there by 40% in these test shops.

Custom AI Models for Each Shop

1What we learned, I think also very good facts to know, to understand, that important is that we had individual model for each product in each shop.

Because if you have a chain of shops,

You can't use one model for every shop because each shop has its own location.

So for example, if you have a shop in a city center and you have a shop somewhere in the suburbs, they have very different selling behavior.

So people in the city center, they buy some more snacks or more things only for lunch.

And if you have shops in suburbs, most of them sell some family, people come.

and buy for family, they buy different sortiment of meat.

Also, if you have some public holidays, it has also influence on your selling behavior.

And different public holidays have also different influence.

shops differ in their trends.

Also, if you have some similar products, it can compensate for, for example, sold already one product in your shop and you have a similar product, it would compensate for this product.

This is also important to consider.

And when you detect some anomalies, because it's always important like in AI projects that you look if you have maybe some anomalies inside and you need to filter them out.

And for that you need at least data from three shops to do this.

Also, the weather influences demand.

This is also important to consider.

So if you do such eye models, you need to keep to weather forecasts because if there is a good weather, like people grill and they would buy more meat if the weather is bad.

So the sales would be different.

a fun fact was that because this meat chain in fact belongs to a pork producer and that's why they were selling a lot of pork products inside but what we noticed that they could sell much more of

much more of chicken meat if they had it in their assortment.

This was not the main goal of the project because the main goal was to reduce waste but at the same time we improved their product matrix because they had some chicken meat in their assortment but very low choice and then we offered them to increase this matrix and at the end they had like better sales even.

Conclusion

This is, in fact, everything I wanted to show you today on practical cases.

I hope it was interesting and, in fact, showed you the variety where we can use AI.

This was my main goal, in fact.

If you have any questions, I'll be happy to answer them now or later on or via email.