# Transforming Retail Through AI: Inside ALDO's AI Evolution with Fatih Nayebi Episode 16 of AI with Sach, a podcast hosted by Sacha Gucciardo in Montréal, features Fatih Nayebi, vice President of Data and AI at Aldo Group and lecturer at McGill University's Desautels Faculty of Management. Aldo Group's VP of Data and AI explains how a traditional retailer builds AI teams, forecasting systems, generative AI tools and data foundations, and measures their return. - Show: AI with Sach, hosted by Sacha Gucciardo in Montréal (https://aiwithsach.com/) - Episode: 16 - Published: 2024-11-18 - Length: 30 minutes - Language: English - Page: https://aiwithsach.com/episodes/16-transforming-retail-through-ai-inside-aldo-s-ai-ev/ - Video: https://www.youtube.com/watch?v=dlPQdr86RTo ## Guests - Fatih Nayebi: Vice President of Data and AI at Aldo Group and lecturer at McGill University's Desautels Faculty of Management (https://aiwithsach.com/guests/fatih-nayebi/) ## Summary Sach talks with Fatih Nayebi, Vice President of Data and AI at Aldo Group and a lecturer at McGill, about turning a traditional retailer into an AI-powered organization. Fatih describes how his office is organized: data governance, a data platform team, data product teams and a science team, with product managers from each business unit defining the problems. He walks through the revenue growth management program: demand forecasting over a 26-week season, markdown optimization and order fulfillment optimization, built with Evada Labs and partly funded by a Scale AI grant. On generative AI, he covers product and SEO descriptions, an internal RAG assistant for store employees, and a conversational, trend-aware search proof of concept from a hackathon at McGill. The conversation also covers the data problems that slow retail AI, privacy rules such as Law 25 and GDPR, a lakehouse architecture on AWS, how ROI is measured at four levels, and how projects are prioritized. Fatih closes with advice for newcomers: respect data quality, expect experiments to fail, and have people who can explain honestly what AI can and cannot do. ## Key takeaways 1. Sort each problem first: does it need intelligence (a brain problem) or just an application or process fix (a muscle problem)? 2. AI projects stall on data: models need many years of clean, complete, consistent data, not just data good enough for daily decisions. 3. Measure ROI at four levels: bottom-line impact, operational metrics, model metrics, and adoption and usage of the product. 4. Not every experiment reaches production; sometimes the data is missing or expectations are far from results, and teams should plan for that. 5. Agile development alone doesn't cover data practices; data quality, observability and governance have to be built in deliberately. 6. Keep people who can bridge technical and business teams and explain plainly what AI can and cannot do. ## Verified quotes Each quote is checked word for word against the recording. > So whatever we have a problem that we want to solve, we look at it from if it's a brain problem, intelligence problem, or is it just a muscle or just application or process problem? > — Fatih Nayebi, AI with Sach episode 16, at 2:02 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=122s) > The prescriptive and predictive models are going to be still very important because the gen AI, of course, is not there to predict things, but more so to generate things. > — Fatih Nayebi, AI with Sach episode 16, at 26:20 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=1580s) > Many of us, we come from the engineering background with application development experience or data engineering type of background. And sometimes we neglect the importance of the data itself. If we do agile development, agile by itself is not going to take care of the practices around the data. So you need to infuse that. > — Fatih Nayebi, AI with Sach episode 16, at 27:46 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=1666s) > We need to experiment things. We need to try things. And not everything that we do is going to go to production because sometimes we find out we don't have the data or the expectations are far from what we're getting. > — Fatih Nayebi, AI with Sach episode 16, at 28:23 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=1703s) > To ensure that you have the right approach, you need people that they can bridge the gap between the technicality and the business side of it so that they can transparently explain what can be done and what cannot be done. > — Fatih Nayebi, AI with Sach episode 16, at 28:34 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=1714s) > So the key idea was to make the search purely conversational and trend-aware. So instead of just you go to website and do some keyword search, you could talk to the website and then explain what are the type of products that you're looking for. > — Fatih Nayebi, AI with Sach episode 16, at 9:33 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=573s) > So, one of the challenges with Gen.ai today is it's not physically aware. And there are going to be physics-oriented machine learning models that they can understand physicality, dimensions, volumes, the capacities, and so forth to be able to match things accordingly. > — Fatih Nayebi, AI with Sach episode 16, at 11:47 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=707s) > The data sometimes is not managed accordingly to become a data for AI. So, in most of the cases, you have the current data that is well managed and has a high quality to do day-to-day decisions or maybe near, you know, past, near future type of decisions. But once you want to make machine learning models, you want to have many years of data that are very quality. > — Fatih Nayebi, AI with Sach episode 16, at 14:19 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=859s) > One thing for sure is you want to be transparent. You want to showcase how you collect the data, why you collect the data, and what would be for the client in return. > — Fatih Nayebi, AI with Sach episode 16, at 16:48 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=1008s) > For machine learning, once you have this established, you don't need ID type of information, right? You want to have a data that is at the aggregate level, can be used for modeling. So, you want the data to be anonymized such that it does not include any identifiable information and then use that for the modeling. > — Fatih Nayebi, AI with Sach episode 16, at 17:12 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=1032s) > I see AI to become a UI between the human and the computer, like from the human computer interaction point of view. To simplify many things, we've been doing programming or we've been trying to create all these user interfaces, always to be able to make computers work for us. > — Fatih Nayebi, AI with Sach episode 16, at 25:05 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=1505s) > We needed to implement first the fine-tuning, but then later on to implement, you know, with pine cone, a RAG from the scratch. And then finally we moved to the OpenAI's assistance API, and now that's behind the scene. > — Fatih Nayebi, AI with Sach episode 16, at 8:56 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=536s) > So, at the end of the day, whatever you do, you want this to have an impact on the bottom line, either through profit or we want to measure the ROI for each of the products that we build. But then, to be able to get there, you always have a sort of operational metric. > — Fatih Nayebi, AI with Sach episode 16, at 21:02 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=1262s) > The examples I provided to you, like revenue growth management approaches are like, it's just the bread and the butter in the retail. And these are the ones that are the most impactful ones. > — Fatih Nayebi, AI with Sach episode 16, at 22:25 (https://www.youtube.com/watch?v=dlPQdr86RTo&t=1345s) ## Transcript Automatic transcript; speaker labels are automatic. [0:00](https://www.youtube.com/watch?v=dlPQdr86RTo&t=0s) **Sacha:** Welcome back to AI with Sach. I'm so, so pleased to welcome Fatih with me today. Fatih is not only the Vice President of Data and AI at Aldo Group, he also serves as a lecturer in the Master of Management Analytics program at McGill's Desautels, teaching advanced courses in data science and machine learning. Fatih has led numerous digital transformation initiatives and established multiple centers of excellence. His track records include developing over 70 mobile apps, deploying 12 AI products and creating scalable data systems. I'm really, really honored to have you with me today Fatih, welcome. [0:41](https://www.youtube.com/watch?v=dlPQdr86RTo&t=41s) **Guest:** Thank you very much for inviting me. I'm really happy to be here and looking forward to our great discussion. [0:47](https://www.youtube.com/watch?v=dlPQdr86RTo&t=47s) **Sacha:** I want to start with a question for you. Since you're the VP of Data and AI at Aldo Group, what's your vision for transforming a traditional retail company? an AI powered organization? [1:03](https://www.youtube.com/watch?v=dlPQdr86RTo&t=63s) **Guest:** Thank you very much for the question. We want to enable AI and infuse it into every aspect of Daldo group, not just such thing that maybe maximizes or minimizes something, but a technology that can help every aspect, help the humans, but also help to automate the processes and help to bring insights. When we look from the AI point of view, there are different types of AI, such as the Gen AI or the Prescriptive or the Predictive AI. And in the retail, in the traditional retail, we have many different types of problems that we want to deal with. There are aspects on the pricing, on the designs of the products. Aspects on the workforce management, on the demand forecasting and so forth. So whatever we have a problem that we want to solve, we look at it from if it's a brain problem, intelligence problem, or is it just a muscle or just application or process problem? [2:14](https://www.youtube.com/watch?v=dlPQdr86RTo&t=134s) And we try to identify the opportunities to enhance and make things more intelligent with AI. And [2:21](https://www.youtube.com/watch?v=dlPQdr86RTo&t=141s) **Sacha:** we know that a strong team structure is very important to do just that. Could you walk us through your AI team structure at Aldo? How do you organize your teams across data governance, analytics, AI in general? Yes, [2:38](https://www.youtube.com/watch?v=dlPQdr86RTo&t=158s) **Guest:** thank you very much. So everything starts with the right problem and the definition of that. We have product managers from different business units that they're the ones that help us to define the problem and then find out what are the budgets for it and how we can actually solve this type of things. Then within the organization of the data and AI, we have data governance team that deals with the data quality, data management, master data management as well as the data privacy aspects. We want to ensure that we have good data which is quality and is accessed accordingly and we follow all the regulations accordingly as well. Then we have a data analytics platform that consists of the data architects and the data engineers. They bring in the data from multiple sources into AWS. We use native AWS technologies to do so. We have a data analytics products team. [3:41](https://www.youtube.com/watch?v=dlPQdr86RTo&t=221s) Everything we do is going to be product-based. We want to bring value with every possible insight, data set, or the model. And the data product team is going to have the scrum masters as well as the product owners and the data analysts. And finally, we have the science team, the AI team that consists of machine learning engineers and the data scientists. So the key thing is that you make them work all together in a collaborative mode to achieve very good results because, as you know, AI is really multifaceted. You need to deal with many different aspects from the scalability, architecture of the data, but also access the data itself, the quality of that data, and ensure that whatever you build is really reproducible. And we do that through having segregated accountabilities, but also making things accountable under the office that I'm managing. [4:38](https://www.youtube.com/watch?v=dlPQdr86RTo&t=278s) **Sacha:** Great. I'd love to hear about the AI-powered revenue growth management that you developed at Aldo. What kind of business problem does it solve? And what were the key challenges in implementing it? [4:52](https://www.youtube.com/watch?v=dlPQdr86RTo&t=292s) **Guest:** Yes. So when I started at the Aldo group, I was looking at AI opportunities. What exactly can be done and how we can augment and make processes a little more intelligent, let's say, and we did analysis with the product managers in terms of the opportunities. And at the end, we identified use cases such as demand forecasting, markdown optimization, and order fulfillment optimization as very important ones to start with. Eventually, we wanted to go for assortment planning and optimization as well as allocation, but that would require longer time horizon, time series forecasting. And before getting in there, we wanted to really focus on the in-season type of use cases. In retail, we have seasons. We have two seasons. Each of them are 26 weeks. With the demand forecasting, we forecast the number of products that are going to be sold and demanded at different stores across North America, United States, and Canada for 26 weeks. [6:00](https://www.youtube.com/watch?v=dlPQdr86RTo&t=360s) And then use this information to do mathematical optimization for the order fulfillment. So the model decides from where a product that was requested from website or the e-com should be shipped. And the markdown optimization optimizes the levels of the markdowns across different regions with that demand forecasting as well across multiple different constraints. Eventually, we want to go and do demand forecasting 10, 12 months ahead of the season starts and use that to do assortment and allocation optimizations. But the revenue management was planned such that we managed the in-season first. We teamed up with Evada Labs and we went through a round of funding from the Scale AI. So the project has a grant and up to 40% of the cost is reimbursed. And within that process, we are working on implementation of the POCs and the MVPs. [7:05](https://www.youtube.com/watch?v=dlPQdr86RTo&t=425s) So far, we have implemented POCs for demand forecasting as well as markdown optimization and started to do the POC for the order fulfillment optimization. And we finalized the MVP. We are about to finalize the MVP for the demand forecasting and working on the MVP for the markdown optimization. So [7:27](https://www.youtube.com/watch?v=dlPQdr86RTo&t=447s) **Sacha:** much for sharing how you're leveraging predictive AI. Now, I'm curious, how are you leveraging generative AI at Aldo? Are there specific use cases in retail, customer service, or internal operations that you're using there? [7:43](https://www.youtube.com/watch?v=dlPQdr86RTo&t=463s) **Guest:** Yeah. So, like, everyone wants to talk about Gen AI. Of course, it's an interesting and very great technology. We have multiple applications of the Gen AI. For instance, we use image generation type of techniques to have inspirations for store design and product design. And then we're using large language models based on OpenAI's ChatGPT assistance API to write product descriptions as well as SEO descriptions. We also implemented our own assistant that is also based on OpenAI's assistance API. It's a RAG model. We implemented this assistance with additions of the specific documents for our store employees so that they can talk with the chatbot about some of the technologies that we have, but also get better in terms of, you know, sales and the marketing aspects. [8:47](https://www.youtube.com/watch?v=dlPQdr86RTo&t=527s) It's open to everyone, but we went through multiple stages of this implementation because this is really a bleeding edge. We needed to implement first the fine-tuning, but then later on to implement, you know, with pine cone, a RAG from the scratch. And then finally we moved to the OpenAI's assistance API, and now that's behind the scene. We are also doing some of the research projects across universities. One of the key ones was coming, that it came from our hackathon, retail Gen AI hackathon that we organized at McGee University. And that one, we did a POC for it. So the key idea was to make the search purely conversational and trend-aware. So instead of just you go to website and do some keyword search, you could talk to the website and then explain what are the type of products that you're looking for. [9:50](https://www.youtube.com/watch?v=dlPQdr86RTo&t=590s) You know, you could say, I want to go to this type of occasion, and I'm looking for this type of style, or I'm going to be looking like that specific type of person. And this would be able to showcase you some of the look based on image generation. You could say, okay, I like this, or I don't like it. And then personalized recommendations based on your picks. So this is another example. We did a POC. We are in the stage to look at MVP, to plan for the MVP. [10:21](https://www.youtube.com/watch?v=dlPQdr86RTo&t=621s) **Sacha:** Really, really cool. So I'm very curious, and I'm going to throw a question at you. I've heard recently about Ambient, I think it's Ambient AI or Ambient Technology, where you basically utilize sensors and IoT to really get a sense of, for example, in your case, in retail, it would be your stock, for example. Is there any plan to get that on the roadmap for you? [10:45](https://www.youtube.com/watch?v=dlPQdr86RTo&t=645s) **Guest:** To be able to visualize and then to change the way that we manage the stocks, you mean the inventory? That's it, exactly. Yeah. So, like, right now, we're using the computer vision technologies to just to ensure that we have right traffic at the store, and then we can capture it. This is all privacy aware. And we were looking at also some fit technologies to model the, you know, the fit, and then based on that, to see how the products are going to fit. We have done also a machine learning model to anticipate the troubles with the fit, with the comfort. And then that informed some of the insights for our designers to increase and enhance the fit and comfort of some of the products. We are also looking at a research project that we want to generate the images in the environment and based on the preferences and place outdoor products accordingly on top of them. [11:47](https://www.youtube.com/watch?v=dlPQdr86RTo&t=707s) So, one of the challenges with Gen.ai today is it's not physically aware. And there are going to be physics-oriented machine learning models that they can understand physicality, dimensions, volumes, the capacities, and so forth to be able to match things accordingly. And some of the research is done today is to really to do that. For instance, if you want to design a store with, like, an image generation type of model, because it doesn't know all the perspectives and the physics of it, it's going to generate something. But it may not be in the right proportions that you're looking for. So, with [12:33](https://www.youtube.com/watch?v=dlPQdr86RTo&t=753s) **Sacha:** Aldo's global retail presence, how do you approach scaling AI solutions across different markets and regions? [12:41](https://www.youtube.com/watch?v=dlPQdr86RTo&t=761s) **Guest:** Very good question. We have the Aldo Nord America that owns Aldo stores in the United States and Canada. And then the Aldo Group International that has a franchise system across the globe. So, whatever we do, we want to ensure that we do it in Canada and United States first and actually works. And then we can customize it for the rest of, you know, other countries in the globe. The trends, the preferences, the way that we manage the stores, even the size runs of the products, you know, how the sizes are distributed are going to be different across different countries. And we want to ensure that we do the right merchandising aspect for the United States and Canada. And once this model is good enough, then we can go to the rest of countries. But, of course, you need to think about all the scalability aspects as well. So, whatever model that we build is run on top of scalable, you know, serverless type of approaches on AWS. [13:49](https://www.youtube.com/watch?v=dlPQdr86RTo&t=829s) And it can be easily expanded to the new use cases or to the new regions in the world. I [13:56](https://www.youtube.com/watch?v=dlPQdr86RTo&t=836s) **Sacha:** want to go back a little bit to the use cases of inventory management, demand forecasting, etc. What are some unique challenges in applying AI to these retail problems? [14:09](https://www.youtube.com/watch?v=dlPQdr86RTo&t=849s) **Guest:** Unique challenges or maybe not so unique challenges is always about the data itself. And not just for the retail, but I think for every industry. The data sometimes is not managed accordingly to become a data for AI. So, in most of the cases, you have the current data that is well managed and has a high quality to do day-to-day decisions or maybe near, you know, past, near future type of decisions. But once you want to make machine learning models, you want to have many years of data that are very quality. So, that is one thing when we started to think about all of these different types of models that we want to build. We did POCs and we tried to anticipate the problems that we had with the data and then to go find the data. There's going to be challenges with the incompleteness and maybe duplications or inconsistencies to fix those type of data issues first. [15:10](https://www.youtube.com/watch?v=dlPQdr86RTo&t=910s) And then start to sample the data to see how and do we have enough of the predictive power to achieve our, you know, expectations, the business expectations for each of those predictions. So, those are the typical challenges. The other challenge is that you want to be super, super efficient and not burn a lot of compute. Instead of just bringing all the data and try to run clusters and the GPUs and that is going to cost you a lot, you want to achieve the most with the less GPU. And we have a lot of data. Like every retailer has a lot of data. You have like sensors data, you have the transactions and the sales. You have every five minute, ten minute event, like eventful type of data. So, how do you process them and bring them, aggregate them to the level that is used, that can be used for AI as something that's required of computation power. But now you want to be conscious about the cost that is going to incur. [16:12](https://www.youtube.com/watch?v=dlPQdr86RTo&t=972s) **Sacha:** That's a great point. You were talking about personalization a little bit before that. How do you balance personalization through AI with customer privacy concerns in the retail groups? Yes. [16:26](https://www.youtube.com/watch?v=dlPQdr86RTo&t=986s) **Guest:** Yes. So, there are many new regulations around the privacy across United States. At 12 states, they introduce different but similar type of privacy regulations. We have also the Law 25 in Quebec and there's the GDPR. There are quite a few important aspects to cover. One thing for sure is you want to be transparent. You want to showcase how you collect the data, why you collect the data, and what would be for the client in return. And once you provide and communicate all of this through right consent management, then you want to ensure that you have the right retention and disposal mechanism. So, how long are you going to keep the data and then how do you dispose it? For machine learning, once you have this established, you don't need ID type of information, right? You want to have a data that is at the aggregate level, can be used for modeling. So, you want the data to be anonymized such that it does not include any identifiable information and then use that for the modeling. [17:36](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1056s) That's what we do. You [17:38](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1058s) **Sacha:** oversee both data platform and AI initiatives. How do you ensure that your data infrastructure supports advanced AI capabilities? [17:48](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1068s) **Guest:** One of the key things to remember is that you don't want the data to be too aggregated or like in a form for the data warehousing. So, before I joined, there was a good investment on implementing things on Redshift. And Redshift is AWS technology to do like the cloud warehousing. There are quite a few different examples of these technologies, but that was the one that we implemented at Daldor Group. But there's also the data lake. So, now there are different paradigms in terms of the architecture. architecture, if you want to unify AI with the BI type of use cases, you want to have some sort of modeling beforehand to do ETLs. And then you also want to be able to do ELTs such that you don't do too many of the modeling. You require a type of architecture that enables both of these type of use cases. And nowadays, we talk about lake houses that enable the data that can be used for the AI or for the BI. [18:56](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1136s) So, we do everything on the data lake, use AWS S3 technologies. But then we try to structure it as much as possible so that the data can be used for either BI or the AI use cases. So, there are different regions. For instance, you have the raw zone and then you go through a prepared zone and then you have a refined zone. And within refined zone, we build data sets for the corporate, for the enterprise, and then very aggregated data sets for the use cases. An example is that we bring in all the data from the point of sales, but also the order management system, from the e-com, from the SAP, from different places. And each of them, they go through the raw and prepared and then end up accordingly into a refined zone. And now, on top of that refined zone, we need to implement some of the aggregation and the joins and so forth so that we can have a flat data set for one use case for the AI. [20:02](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1202s) But the aggregation method is going to be different for some of the reports and the dashboards. So, we do it differently. We kind of build a data mart on top of the data lake. And this way, many of the practices that we do in terms of the data quality checking, data observability, data management are going to be applied in one place. And BI and AI, they're going to be able to leverage from those practices. That's [20:30](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1230s) **Sacha:** amazing. We know that data and AI has a lot of value when it's done properly. What's your approach to measuring return on investment, on AI investments at Aldo? [20:45](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1245s) **Guest:** We have a measurement approach, product measurement approach as a whole. So, every time that we want to build a product, we want to ensure that this brings the value. And they can have different types of metrics for each of them. So, at the end of the day, whatever you do, you want this to have an impact on the bottom line, either through profit or we want to measure the ROI for each of the products that we build. But then, to be able to get there, you always have a sort of operational metric. So, for instance, can this help to increase the conversion? Or can this help to reduce the markdown? Or to improve the traffic? In the middle, operational metrics. And then, you have the model-related metrics, as you know. If it's whatever that model is, if it's regression, if it's classification, it has its own metrics. And then, finally, you have metrics around the adoption and the usages of the products. [21:47](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1307s) So, if you end up using and providing these models in dashboards, just to show the insights, for instance, we have analytics around the usages of the products that we build. And, yeah. And if we are providing APIs, so do other applications tap into, we also have the log metrics around those requests and the calls as well. And [22:09](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1329s) **Sacha:** how do you prioritize different AI initiatives? [22:14](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1334s) **Guest:** So, we look from the operational point of view, how much of the uplift is going to be, but also from the strategic point of view, how much of the reusability and the capability that we can gain. The examples I provided to you, like revenue growth management approaches are like, it's just the bread and the butter in the retail. And these are the ones that are the most impactful ones. But, at times, we also want to look at how we can make the consumer's lives easier, how we can increase and enhance the customer experience. And sometimes, it's not super easy to calculate and to find out how much of the uplift that's going to be. But, we try to stay positive and then try to find out how the impact is going to be on just operational metrics in that regard. And even just operational metrics, maybe not fully aligned with the dollar values at the end, could help us to decide for which use cases to pick. [23:19](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1399s) So, what we do is, we go through different types of use cases and with different product groups. And for each of them, we try to understand the business value, objective and subjective. And we look at the type of use cases that are required across multiple product groups and the squads. So, they are going to be more strategic. And based on that, we have product managers that they align together. They sit together and try to see what type of insights that they can reuse across the squads. Let's say, because we have multiple product groups, let's say one for the retail and the e-com and the store. So, use cases for those are going to be tackled separately. Another one for the supply chain and the wholesales and the finance. And there are multiple product managers from the finance, from supply chain office, from the retail. And they try to see strategically how we can leverage some of these insights across different places. And every year, we're using Atlassian toolset. [24:25](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1465s) We have something called Giro Discovery. And we draw the line in terms of the value that we can deliver. You know, the value and the budget to deliver the most for everyone. [24:37](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1477s) **Sacha:** Interesting. I'd like to look to the future a little bit. Where do you see the biggest opportunities for AI in retail over like the next three, five years? And how is the Addo Group positioning itself for these changes? [24:53](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1493s) **Guest:** I think the user experience and the customer experience needs to be revamped. I gave you an example of the POC that we made after the Hackathon competition. But that is not implemented yet. I see AI to become a UI between the human and the computer, like from the human computer interaction point of view. To simplify many things, we've been doing programming or we've been trying to create all these user interfaces, always to be able to make computers work for us. Now, we have the bi-directional communication because it can come back and give us clues, but then we can create all of these feedbacks and refine it. Moving forward, I see this aspect to be a lot more interesting and more advanced, such that you can have conversational type of system so that it can understand the context and it has the memory so that you can find what you want from the search perspective, recommendation perspective. [25:58](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1558s) The image generation, video generation, the multi-modality is going to help as well. So maybe you can use it in a sense to do the AR and the VR, like the augmented reality and the virtual reality using the generative AI type of techniques. The prescriptive and predictive models are going to be still very important because the gen AI, of course, is not there to predict things, but more so to generate things. I see opportunities in leveraging deep reinforcement learning for optimization, such that you can learn many different types of patterns instead of just going for the convex or linear type of models instead. And finally, what I really think about how AI can have a big impact is to combine multiple different types of AIs together for the automation, but for better understanding, and then to be able to provide better results to either the customers or even to employees and that they work in the back offices of the retail. [27:08](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1628s) **Sacha:** It's been a very interesting and insightful conversation. I think it's very important to share any last words, any last comments on whatever you want to the audience. [27:17](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1637s) **Guest:** Thank you very much for the invitation. Thank you very much for being here. I'm very happy to share my experience at the Aldo Group. For the ones that are in the beginning of their journey and they heard about the AI, they want to do the AI, I think it's very important to have a multifaceted and holistic view about the type of things that is required. And do not underestimate the importance of the data and how data should be managed and used. Many of us, we come from the engineering background with application development experience or data engineering type of background. And sometimes we neglect the importance of the data itself. If we do agile development, agile by itself is not going to take care of the practices around the data. So you need to infuse that. Also, we need to remember that although that so many of these algorithms and the libraries are so well advanced, but there's still a bit of the research is required to be done. [28:23](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1703s) We need to experiment things. We need to try things. And not everything that we do is going to go to production because sometimes we find out we don't have the data or the expectations are far from what we're getting. To ensure that you have the right approach, you need people that they can bridge the gap between the technicality and the business side of it so that they can transparently explain what can be done and what cannot be done. And once you establish that, then you can have all the partnerships with organizations. And we're very lucky to be in Montreal because there is a great community, academic community, as well as a lot of very good companies that they work on the AI front. So, yeah, so a holistic approach. Data is important. Everything we do is based on data. Experimentation, research and development also are the aspects that needs to be covered. But you still need to do coding and engineering as well. [29:19](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1759s) **Sacha:** Yeah, I couldn't agree more. Data is king and we have a great, great, great ecosystem here in Montreal. And honestly, Fatih, it was a beautiful conversation and I learned so much. Thank you so much for being on today. [29:34](https://www.youtube.com/watch?v=dlPQdr86RTo&t=1774s) **Guest:** Thank you very much. I appreciate it.