Using AI to power a circular economy for aerospace

Good afternoon, waking you up. Uh good afternoon on day four of day one at re invent this afternoon seems to be having a toll on the audience. Um I wanted to start with a question to the audience actually, who has been on Werner's keynote this morning further quite a lot.

So one takeaway that I took from that keynote, apart from architecting frugally and sustainably is a phrase that was on one of the slides that said the worst phrase in English is "we have always done it that way," right? And that brings me to the topic here.

Aerospace industry, aircrafts are in desperate need to attack sustainability and become more sustainable. There are more than 800 aircrafts retired every year. That's a lot of parts and aircraft. And so Capgemini together with AWS has built a solution to track these parts, to inventory these parts, to help the aircraft industry retire, recycle, refurbish these parts.

And we are really blessed to have the business owner and the technical owner of this solution here and they will dive deep on. What's the rationale of building the solution? How has it been built? Technically? How are we using AI to track the parts and there's a bit of Chennai I of course as well.

So over to your Matric, thank you.

So in type of agenda, uh first, we will introduce how the idea came from. Uh we'll explain the solution for sure uh where we are today and also the ambition for the platform.

So to start uh with some figures first to for you to understand uh the idea behind. So i just, you mentioned there, there are of aircraft which are retired every day, every year, every year. So just in terms of number, there is 20% of aircraft today that are older than 20 years old and an air aircraft are retired at 2530 years old, something like that. So basically, there will be plenty of aircraft in the future that will be tired in terms of recyclable. Uh today, there is something like 90% of an aircraft which is recycled, so it's good. But when we come at the part level, there is less part that are remarketed. So basically, there is an opportunity to grow the number of parts which are put back on the solo market. It is what we try to address this, this project and uh and at the same time, so the ambition for that is to try to save tons of co two for sure. So if we dismantle the aircraft in the right way, we think we do something positive. And if we use more parts at the same time, we will avoid tons of raw material extraction. So basically, the idea is to improve the way we are using what is existing today.

So first for everybody to know, you can think that aircraft mountains papers etc are digital. So it's true. But beyond that, there are still papers. So basically, you can't fly an aircraft if you don't have the papers. So it's really important for us and it's where we place the solution.

So, so Claudio will then explain how it is working. But to make it simple, we use AI and machine learning to, to track the records and to be able to, to address the back to birth of parts and to to, to make sure that the traceability issue of parts are addressed and we get everything to be able to use the parts.

So you see the four domain we try to address. So the first one is again reuse. So if we have the traceability, we are able to reuse the part. So it's what we try to do at the same time, the repair. So again, to make the proper repair on an aircraft, you need to know the history to do something good. So thank to the AI module, we are improving the way to repair parts, resale. So again, if we have a possibility, we are able to put part on the market, so we are able to resell it. So it's a new source of revenue for sure. But at the same time, we use something existing and the last one is all.

So we created a solution along with AWS. So on, it started two years ago, the development of the solution started one year ago. So the solution life cycle for optimization I said has been tested. So we will show you the first results. And the ambition for us is to put this solution on the market early 2024.

And uh what is behind? So again, you see on the title is about circular economy. So what we try to promote on this project is to transition from linear to circular economy. So basically linear model is when we destroy value, circular is when you improve and you preserve the value. So it's a way for us to improve how we manage the map of parts and that how we continue to use the parts. So it's ambition encouraging circular economy and Claudio will explain how it is working.

Thanks. So we started as Matthias said, we started about a year ago uh to develop this new platform. And uh what we knew, of course, we knew the business uh uh in camp January, we have a division on that. Uh Mattia knows very well the business i didn't personally, but we knew perfectly the business. We knew some requirements. As you saw, i could have 300,000 documents and pages and we need to analyze in few minutes so hard or maximum an hour, all these documents.

So we knew that we had to do some tremendous scaling in our platform. We knew that of course, we have to use serve because it was the easiest option for us. And we knew that we want to use AI service of AWS. And uh that's an HLD of what we have realized that what we didn't know was what kind of AI and what are, what we, of course, nobody was expecting what such a huge improvement on generative AI that we saw in the last year.

So, but we were expecting something we didn't know. Of course, we are expecting something. So when we started, we tried to create something that was agile. And when i say agile i mean, on both our team is arranged in a agile way we have 1520 people working on that. And uh i arrange it with the regular meetings, uh daily, weekly, something that you probably know. And uh we have an infrastructure too.

We have created an infrastructure, everything here. I mean, this infrastructure is deployed with CDK, everything is deployed using uh CI CD. So everything is ready for any change. And the most important part of where we spend most of the time was step function because we wanted to create an automatic workflow to get the best of the different AI that we want to use right now, we are using 34 different AI services here in this platform.

So the most complex part was to create an automate this workflow on on AWS. And that's what exactly what we did, what we realized. And just to give you an idea here, we have a reactive front end for a web interface, some rest and graph ql API storage open search is the most important thing because we store there all the metadata information that we extract from our documentation cognito plus of course, step function. That's a very high level picture of the step function. If you want more details, scan the qr code that you can find a blog post that we published last week on the AWS APN blog. And to go in details about what exactly we did with Amazon. But that the, the what we wanted to create was something easy and uh to describe the process.

I mean, i remember, i always remember the first call that we did with amazon area, the alan of amazon. And uh and uh they wanted to test our platform, we show the platform they wanted to test. And the question was how can we share with you the data? what are the api that you need to implement? what, what, what do you need to, to, to test the platform? i say just give us file. we don't care about a p of course, we have rest a p but just give us files or zip file. uh put everything in one huge zip file, 20 gigabyte file and give us.

Ok. what's the format that you prefer? what are the folders inside the zip file? again, we don't care, put everything in a zip file, upload on the platform and we will do, we will analyze all the files, we will connect together all the files indication about what you can reuse suggestion and possible problems on the on the documentation that maybe you are missing on that part. that's what we have real.

And to give you an idea of the service that you use for the eye. so you upload the fire nest tree, we unzip all the files. then we do sr with text for all the documents, 300,000 documents on average per plane. then with the output of text structure, we have a sage maker inference and point deployed. of course, we train the model. it's a bird model. we train with our data with the data that our customer share with us. and then with the information that we obtained by the age maker, we have a classification of the document. we highlight the 1520 documents are mandatory to reuse the airplane.

and and for this 1520 documents, we extract the information, how we extract the information. we use some additional official of text structure that basically are forms and queries. we had some custom code and spoiler. we add even the generative as soon as amazon deploy released bedrock in production, we implemented. and right now we have in production bedrock to do some specific data extraction that it works so well and i will have a slide later.

so opensearch is the heart of our system because we store all the information there. and once that we have stored information, we will run some custom script to really understand if you have all the documentation required for, for, for your part. if you have some missing uh missing stuff, if uh uh you can improve in some way.

and here you can see a video of the platform. basically, you just need to upload a file, wait until everything is is executed. and first residents were pretty good. we did an mvp, we, we create an mvp in a very fast way because of the service of amazon, the serve the automation that we have in our experience of the business. and first was, was amy. i mean, we have a lot of data that we collect under the airplane. and basically people are writing with are writing data on papers and we extract something that i barely see. i barely understand when i see with my own eyes. and the platform is just web application pretty here.

you can see the as honestly right now are much better than this because because we want to even to add the possibility to improve automatically in the algorithm. and when improve means that in case of misclassification of the documents, we the user can change the classification and we use this information to retrain our models and improve it.

so that's what we have realized. and we are pretty uh we are pretty proud of that. and uh but of course, the main problem is over the security. we added this line a few weeks ago because of course, during these weeks and during the event, we had a lot of discussion with a lot of customers and AI is something new for most of us and most of us have have doubts about security and we want to give to our customer the best platform, the best security.

so first of all, of course, we start from uh from the basic stuff. so the war architect review, we are, we we a reviewer, we follow architect framework, we did two reviews, one on the capgemini side, one on the amazon side and what we created, we created the platform there. it's fully encrypted and addressed in transit, follow all the guidelines and it's the most secure that you can create on. uh right now on amazon, that's what's fine but what not, not enough, not enough.

because we are talking about a a solution. and as some solution means often multi-tenant even to have a lower cost of the infrastructure. so we decided that every time that we have a new customer. we don't simply do multi-tenant, but we create automatically with a, with an automatic process, a new aws account, we create a new aws infrastructure we deploy on a separate account. this means that even if we have a bag, we won't share any data. we are sure 100% that even for a b, you won't be able to access to the data of the another customer. that's good. that's great for, for, for security.

and because it's serve us, it means that it doesn't cost a lot more. it's on several, you pay for what you use. if you have two accounts, three accounts, 100 account or one single account more or less, the cost is the same. so that's great.

again, not enough, not enough. because as i said, we want to improve our robot and we would like to use the information that you can give us for instance, misclassification or any document to train and improve our model. but some customer say we are scared about that. i'm not sure that you can uh can uh maybe generate some data leak because of that.

so how we are managing right now? basically, we are the two options. one that basically you can say no, i'm not. uh i don't allow data collection for the amazon service because of course, the problem is even on the amazon side. so if the data that we pass, for instance to do text track is not used by amazon to retrain their own model and same stuff on our side. you can choose not to, uh, give us this possibility and turn it off honestly with the announcement that we saw yesterday and the day before and, and i'm referred to, uh, be drug g grades.

yeah, we are, are already working on something better for the future, but for the moment, this, it's fine for our customer, this allowed to have the best security and guarantee to our customer not to share any kind of information with someone else.

so claudio explain you the technical things. the back office was introduced, there was early adopter on the solution. so you see the names and there was saffron, there was a france and there was also amazon primer. so there was the early adopter they accepted to share data with us, particularly to train our model without them training of the robots were impossible. so we thank them again.

and we proposed to look at the first video to it will explain why saffron decided to support the project and the sound is not working but you you have the messages, there was sound audio problem. we have the support team coming on stage to help us. we will fix you. j i works well. it's just a cabling issue.

so in the meantime, again, so we do some tests with saffron. they accepted to provide landing gears information to the landing gear. is the landing gear of the aircraft. saffron is making the maps of the landing gear

So each time you need to do something, the alliance is sending back the to sa sa is not allowed to do that. And there are plenty of customer. So we get all the information of operation of the l years. And what we try to do with them is to get all the information.

So you see on the first video, some of example and retrain the robots to understand how the how the documents are formatted. Where are the important information, how to extract the information. If there is something specific that the mechanics want to just to mention in the information, we need to identify this information.

So this was all the important things that we get from them to learn our modem. And if the sound is not working, i will just explain the video. Uh so what was important for for sa and this use cases?

So first using a i is good for them because there is a productivity savings first. So i will keep the video up. So first there is the productivity savings uh when there, when they are getting the landing gears from the airlines, they get the full set of documentation. So analyze the full set of documentation, taking a lot of time. So first we use a i to save time in analyzing that they are also able to get first document from the airline to make sure that the map. No, they are proposing is the correct one. So it's also a way to improve for them. The map. No, they are offering to their own customer.

And then also during the map north, thanks to this kind of solution, they are also able to identify the part they can reuse maybe on no landing gear. So previously, they were not considering using the part anymore. So thanks to this kind of solution, they can also improve the number of aircraft that can take profit of the parts and improve the reuse rate of the part as i have mentioned at the beginning.

So what you see here are the first results. So again, we take an example. So the orange person is a, is a crimi engineering mechanic. So we have internally within crimi people that are able to inspect their craft to make it simple. So we ask him to analyze the set of documents we get from the airlines. So we ask him, ok, let's let's make a sample on some parts. So we make a sample on 18 parts on the landing gear. You have more than 200 serialized part. So it's just a short sample. And in parallel of that, we ask the robots to do the same. So 2.5 to make the sample from the human and less than 30 minutes to make everything with the robots.

So he get all the document, he sorted everything, he extract the information, he have identified what we are missing. And also he packaged everything for the macs to be performed in the right way. So without the first result, thanks to that, we decided we have to continue exploring further the solution.

We do the same with air force but on a computer aircraft, not only on the landing gear and then we are also now discussing with other customers worldwide that are interested to, to use this function to improve the maps to improve transaction of aircraft. So from an airline to another and also improve the way they are doing procurement of parts.

So when they are making mountains, sometimes you need to change a part. So for them, they are looking on the single markets about the possibilities. So thanks to the solution, they can evaluate more possibility than before. And so for them, it's also a way to improve the way they are doing them.

So what next at the beginning? Uh i said that the solution is still in the development. Uh it will be available in the next weeks within the aws marketplace. So we are going to make an sspo a private offer with uh with aws. So everything will be available.

So what you see on the screen is that it is working there is existing solution. Uh what we call maten information system uh where you you can find all the information uh on the aircraft. In parallel of that, you have the papers. If you get lucky, you have the pdf documents. So we get all those documents. We put everything like claudio explain within the solution. We ask the robot to do the magic and still don't know how it is working, but you do and then we get structural data.

So first we notified where there is issue, we make a recommendation to improve the situation and then we package everything for again the landing gear of the aircraft or whatever the part we are analyzing to be reinstalled on another information system to be to be reused. So it is how it is working and soon available for everybody.

Uh then uh what what next? Uh so what you see first is a i to improve the way we are analyzing. So again, we save time. If we save time, we save possibly parts. So it's, it's our objective, but we, we say it is not enough.

So we decided to look at what the document we are analyzing can bring us. So what we see on the document um other information. So on the map, no documents you can see for sure the reference of the part, the number of light of shoes, of the parts. And you can also find the materials. So you can know that the the part is about titanium aluminum or something like that.

So it's now tricky to pick. So you see all the issue we have on the supply chain to, to, to, to just get the appropriate material to manufacture new parts. So we consider there is also a way to improve globally by just improving the way we use the materials if we cannot use any more the parts.

So the system will let you know, you put the parts inside if we say ok, the part is ok, you have all the papers, it's free to use so you can continue to use. It will let you know, ok, the part is not ok, you have a missing document or something like that. So you have something to, to manage or we will let you know the part of no more value it's over, but it's titanium made. So maybe you can do something.

So with how it is working, so we make some party notification, we notify the whole materials. So we are not inventing some. There is a list of rare materials available. So we just map this list with what we find into the documentation to identify where we have opportunities and then we, we are using what we call.

So just to identify where we have the specific uh rare materials and we classify again everything. So if you are a user of the platform, you can look and say, ok, i want first to see the part, we still have a value, the part i can use on an aircraft. So he's sorting, then you said, ok, i, i don't mind. I just want to see where i found the aluminum and then the platform is again, sorted.

So it's a, it's a way for us first to improve reus it parts again. And if we are not able to use the parts, it's a way to improve the recyclable and to continue to avoid ho and uh as all the presentation of the week, uh we present ja i, so it's mandatory, a concrete example of how it is working.

It's mandatory to have generative a i. Of course, we cannot. But first point, this is a real generative a i implemented in production right now, our customers are using on our platform. So that, that's something that is and it is bringing uh business value because we, as soon as amazon released bedrock in uh in frankfurt about three weeks ago, we implemented, we did some investigation. Of course. Thanks even to amazon and the support that always give us.

We did some investigation. We tried to find out the ex case where we can find business value. And honestly, we found out four of them, we implemented only the first word we wanted to wait for las vegas re invent. We don't want to re invent the wheel and, and we were expecting some service and that we are going to use in the next weeks. Of course.

So if we implemented this use case, let me explain a little. So first of all look at step function. So that's the real complexity that we have implemented those step function. And as you can see, there are step function inside another step function inside another step function to make, make it more complex. But this give you freedom. This gi it's the aid that i was saying before.

Basically that that's something that allow you to add another step inside the workflow to use a new a i service. That right now it's bedrock cloud v two. So we are using this case, cloud v include virtue, but of course, could be tightened and now is in g a could be another service that we arrive in the next month. It is not a problem, we can easily implement.

So what's the use case? Look at the, the, the picture, there's, that's a form one, that's a mandatory document. If you want to reuse a part of the airplane and how you can maximize the part of the airplane. You can see that this as a form one.

So the european regulatory and let's say that i want to reuse this part of the specific part of the airplane, not in europe and the us. So if you want to do it, you can do it, but you need to double check that you have the flag, this tiny flag that the cell is 12 c and you have a sentence in the note but nothing is so simple. It's not a specific sentence. It's a general sentence that we found out that it is different between islands and even inside the island. And the same is it different? It could be even in different languages.

So for these generative is perfect. So we create, we tested first on the console with the help of the amazon. We we did some prompt engineering. We find out different sentences that we can send to cloud to extract this information. And then we use in our pipeline, we implement inside our step function pipeline inside the land. Of course, this additional step in our work flow.

And we highlight to the user that it can be used in different country or even that you have a problem because we found out that sometimes there is a mistake on the documentation and there is the flag but there is no sentence or there is the sentence but there is no flag and that's a problem of course. And we implemented pretty fast just to try to understand with me uh full time equivalent. Ah how many days do you think we take? I'm not thinking just to bedrock implementation. I think you do the whole implementation right in the land, implement the the bedrock kpi test, different different prompts, do some automatic test and do some final test validate and bring in production five days, 10 days, 20 days. Try to guess three months.

I'm sorry was there is three weeks ago bedrock in in uh in, in frankfurt one day, five days, five days because of course, be network is great, great service ready to be used. It's wonderful that you go on the console. We, you basically, you, i can manually copy and paste what is the output of, of text on, on the console and try to understand if it works and which one of the different algorithm can be used. But even because we have created a fully automated platform based on the step function uh deploy with cd k and c i cd.

So even on our side, we are ready to plug new service and that's what we did. And that's the reason that we took just five days and then we implemented just one because of course, we don't have enough time. Las vegas was coming. We didn't want to re invent the wheel. In this case, we are not using the knowledge base.

We are expecting something on the knowledge base during las vegas. So we wanted to wait for the other use case that required the knowledge base and we didn't want to manage the infrastructure to, to, to do the knowledge base. So that's, that's the beauty, that's a real implementation that brings you value and improve the usability of part of the airplane.

Oh yeah. Concrete example. Yes. So as claudio mentioned, so thanks to just the generative a i, we are here able to just say you can not only reuse the part in europe, you can use it also in the us. So basically, it's improving the way to reuse the parts without generative a i, we make, we make also the analysis on using standard a i i would say like that. So the results were not positive.

So what we do on the platform is just identifying the proper use case where we apply because there is a cost down behind. So we try to limit the cost as well to do something good again to improve circular economy. And how much does it cost? What do you think? How much does it cost to do this type of analysis? On average? Of course, because we have different document, different formats and different sites give you a number a rough number. Of course, it depends on the use because we are using the most expensive version of cloth vu so room for improvement about 10 cents, a little bit less than 10 cents fine acceptable. It's not a problem.

Oh, if i need to apply this to 300,000 documents, it's not honestly, i mean, we cannot spend so much money. So the the sorry, sorry to interrupt you with you. The point that i just want to share is you need to use the right at the right moment to create a good pipeline and be sure that you are using the best even in terms of cost. But of course, in terms of performance, precision and stuff like that. And that's what we, why we have created this platform and we will end up.

So i would say it's not just an mvp, we are not just trying to do something. So now it's becoming a real as we have, we have internally. So it's a way to digitalize the airs to use the latest technology to do something positive. It is supported by our ceo and it's uh something we continue to develop in collaboration with aws. And uh and we hope that we will have plenty of customer using that uh to, to promote circularity globally.

And if you have any question, we are there with claudio and to, to answer.