№8 – AI-Powered Perfumery with Milan Pavlovic

In this episode, we explore the intersection of fragrance and artificial intelligence with Milan Pavlovic, a software engineer and fragrance enthusiast who founded Scentalytics. This startup has garnered attention in recent months for its work in utilizing data analytics and AI to redefine the norms of the fragrance industry.

EPISODE LINKS:

• Try AI Perfume Creator
• Scentalytics Website
• Milan’s Profile at LinkedIn
• Scentalytics’ Profile at LinkedIn

Ever dreamt about going to space? Connect with Tanya!

• https://www.instagram.com/neparfumer/
• https://www.coachmironova.com/

Transcript

0:00

And then I can say, okay, one perfumer is pretty suitable for that brand and so and vice versa. So I can match them based on those scent profiles. oh Welcome to Not a Perfumery podcast, where we talk about science in relation to wellness, art and innovation. My name is Tanya Mironova. I'm an factory art educator and your Denmark based host. My guest for the today's episode is Milan Pavlovic, a software engineer and fragrance enthusiast who specializes in artificial intelligence and data science. He owns Centalytics, a analytics and AI startup. focused on fragrance industry. Milan, you're welcome. Thank you. It's nice to be here. I don't want some of our listeners to fall asleep when they hear data science and AI in one sentence in a podcast which is related to sense. please, Milan, tell us in a few words what is that you're working on. I'm really passionate about data and fragrances as well. So when I started to gain interest in fragrances, this was around two years ago, two and a half years ago, I noticed that there are fragrances contain a lot of data. They can be described with a lot of different kinds of data. Since I am coming from the software engineering and data science uh background, I started to notice this data, which I'm seeing like when I'm reading about perfumes, I see that... They have seasonal suitability, uh occasional suitability. They have different notes and accords and so on. So all this is inspired me to start thinking, uh okay, so if I collect enough of that data, would I be able to train some AI models which could help us to predict some characteristics of future? fragrance products before they are created. So basically that's what I'm doing now in Centalytics. I'm training different AI models which could tell us uh how could fragrances look like, smell like, feel like before they are created based on some inputs like desired uh notes. which will be used in the fragrance. I would like to do this in the opposite direction as well. So if we know what characteristics uh should our future fragrance have, we could say them, enter them to the model. So the model could suggest us, for example, to suggest perfumers which notes and aromas could they use to create such a fragrance? And in this way, they could maybe be uh augmented with this initial uh knowledge from my models. They could quickly start uh making their creations. Of course, my goal is not to replace perfumers with the AI, not at all. I just want to help them to be even more creative. and to use their time in more creative way. ah I indeed already have had an experience with a friend of mine. I sent her one of your, how it's better to say, one of the parts of the application, the AI perfume creator. Because technically if you have those models trained, and I don't think we should go very deep into that, but if you have those models trained, you could use them for different applications, right? Yeah, and this AI perfume creator uh actually leverages five uh or six, actually six AI models, which I trained. And it's more, you should use it for like demonstrational purposes of what those models can do. You can enter the notes there. ah And then you can see the characteristics of those perfume compositions, like the accords, uh then perfume type, then gender, weather and daytime suitability. And yeah, this is actually the models I am talking about are those in this application. Yeah, yeah. So, you know my friend, the one who I mentioned, she tried your application. She's right now, we're recording this episode before the exams, so basically everyone is going to present something new during the exams, etc. And she had a new formula in mind, of course, in her Excel sheet. The thing is that she tried the new scent in your app. It showed two perfumes which were 90 plus percent similar to her formula. And so she felt like wow at first, because like she could technically see who used it like that. But to us, I mean like maybe to you, to her, to other users, it's clear that there are some good things about the possibility to just enter your formula and see in which perfumes it used. And there are bad things about knowing this information. If I may, I would like to firstly, uh before we go into the pluses of oh using AI to ask how difficult it was to find enough data in the open sources. This is actually the most important part when you are creating AI models. It's not about... I mean, to me it's nice to train them, to tweak them. uh But actually the most important part is to find the good data and to uh prepare the data for the models. ah For now, I can say that I use data from publicly available web sources. ah I collected the data and I use the data to train those models. So... ah Yeah, it's uh the data is available on sites like parfumo.com or fragrantica.com, but it's not so easy to collect the data and sometimes that's not even allowed to do. Yeah. So I mean, this is kind of my let's say business secret for now, but this was actually the most challenging part to collect the data. But yeah, I use the publicly available data. I found a way of collecting it and yeah, for now that's what I can say. yeah, I mean, the data uh is available, but maybe... people which have more technical background like me can get the data more easily. for you. Well, so this part is clear. Returning to my friend's situation. So a perfumer sees that her formula is like 92, 90 something percent similar to a number of perfumes. What does it mean to her based on the product you built? So you're talking about the possibility of uh seeing the similar fragrances. Yeah. So when you click on that button to show you this similar fragrances, basically I'm turning all fragrances into vectors. That's a more mathematical term. So vector is let's say a list of numbers. I do this on based on notes you entered and I am calculating the... the similarity scores, which you can see with the one relatively simple mathematical technique, which can tell you how similar two things are when they are expressed as those vectors, as those lists of numbers. And if you have a high similarity, this means that your composition, your notes and accords are semantically similar to those products, mean to those notes and records from the products which are on the list. Yeah, so this means that for example if you create, I don't know, some kind of light summer citrus based fragrance uh with bergamot, notes of bergamot, citruses, I don't know, lemon and so on, you will probably see the fragrances which smell in a similar way. ah Maybe not exactly. I mean, you cannot always have 100 % similarity unless you pick exactly the same notes from those fragrances. But you can just see the direction in which your scent will go in terms of semantics of notes and the chords used. by the semantic we mean here that the people's perception. Yeah, I mean uh basically the perception of the scent and scent type. What we receive, the result, means that semantically, as we say, like from the sense perspective, it's, so in case of with my friend, like 90-something percent, similar to most of the people perceive this. Yes, I think that this perception is a good word to say, because probably the scents won't be the same. I mean, you cannot create... Even if you have 100 % match, you won't have the same scent. It's not that accurate. But in some kind of perception, as you said, of a perfume type or uh dominating uh accords, there will be a similarity. I see, I see. Because for the 100 % accuracy of this model, which is totally achievable, we need to have formulae from all the perfume houses, all the perfumes. Yeah, and if I can just add those percentages you see for for for records and other characteristics, those are just estimates from the models. Those are not the exact formulas which have to be used. Maybe it's important to notice this that I don't have I didn't have the data about exact formulations and aromas. And yeah, because This is just the those are just the predictions and the predictions can be expressed as uh Probabilities so if and this is those numbers are also post processed so they fit a uh Rounded chart by chart in in some they are they sum up to hundred percent But if you for example if you see ten percent of powdery occurred this means that there is some powdery occurred there And okay, you need to compare this amount of it with other numbers to see the relationships between possible dominating accords and so on. But this is not the exact formula you have to use. uh this is more like just for the comparison purposes of which accords could be more, I mean, which could dominate more or less. So, yeah. So those are not ex- and the same applies for other charts. If you see, for example, that 60 % is for the unisex gender, this means that this is probably a unisex cent. For example, if after this, if it has 20 % on the male side, then this could be interpreted as it would probably lean to females. But, um... Those are just the directions, not the exact rules. I see. And probably by this point, listeners are just thinking, well, why it's hard to have data? Why it's hard to have exact formula? Maybe you could just comment because I know why, but oh for the listeners, because I you would be the happiest engineer if you had all the data. yeah. So the dream of every data scientist is to have as much high quality data as possible, but that's actually never the case in the real scenarios. I was actually surprised that even, mean, uh those web sources, which I used actually had solid amount of data for, let's say those basic purposes. for fragrance users which explore the notes and and the records. Because I think that this data can be... I mean the fragrance producers actually give... they share the data with others to describe their products. So from this perspective it's understandable that this data exists and regarding the formulas this is probably Those are probably their business secrets and they do not share this because they don't want to be copied by others and so on. They did it with other methods, not data science. Yeah, I mean I look on this through the data science perspective So I saw that some people do share like they call them open source formulas when where they actually Write a list of exact ingredients and the amounts of them used in some perfumes But I don't know this kind of data isn't you can from my perspective you cannot find it so easily as the one I used because I used mostly the data from product descriptions. yeah, from that perspective, it's understandable to me that we have a lot of, mean, let's say, I'm saying a lot, ah but I don't know what this means to other, what can this mean to other people, but to me it's actually a lot because I think there are not so many places on the internet where you can find more. So yeah. One of the things which I believe very special about what you do is that you A, collect the data, B, made it little bit already available for everyone. I mean the AI perfume creator tool. But I believe you have, uh like this is only part of what you do, but Centalytics, other products and other ideas of how to use these data, apart from the publicly available, but more customized ones. What are those? I could create scent profiles for different brands and even perfumers. ah And uh I could also do, so I could maybe suggest uh which perfumer is more suitable for which brand to make fragrances. if you have, so these vectors I was talking about, they enable us to do uh different combinations. ah And to compare, for example, uh I uh have vectors for each fragrance from my database. And I know which fragrance was produced by which uh producer or brand actually. So I can create a scent profile for each brand by averaging those vectors together to create a single vector which represents a brand. But I can also do that for a perfumer. by using vectors from fragrances a perfumer created. And then I can do the same thing like I have show similar fragrances in the AI perfume creator app. So there I'm comparing your vector from your notes with all other fragrances from the database. But in the same way, I could compare vectors from perfumers and from brand. And then I can say, okay, one perfumer is, uh pretty suitable for that brand and so and vice versa. So I could match them ah based on those scent profiles. If I understand correctly, the thing is that each scent is a vector, right? And like if you look at the perfumer's profile, do you also see a number of vectors based on the scents he created? Yeah, I can see fragrance scents and fragrances he or she created and then I have, if I have, ah if I'm able to create the vectors of those fragrances then I can see multiple vectors for a single perfumer for each perfume. And then I can create a single one or couple of ones based on the whole set. Seems like I need to revisit my data science course, which I've given up some time ago, but does it so that now you have, um, like a sort of a database of the most, of the most prominent probably perfumers? and, um, have you like been looking... at this portfolio. I understand that it doesn't look like I'm showing it, it's not probably this, it's that or something, but have there been perfumers whose works are in in all vectors very different? Yeah, some vectors could be very, uh let's say very different from uh some typical ones, let's say. And yeah, this is, mean, we could calculate the similarity match with others, but there's one drawback of that method that if you use a lot of different vectors to create a single one, then it's questionable. What did you get as the end result? Maybe this result um cannot be applied in that way. Yeah, maybe it cannot be applied in that way that I think it can because we mixed up too many different things in a single thing. And then we have something what's maybe not applicable to that use case. yeah. This vector thing is uh all this started in uh natural language processing where words were represented as vectors in the computer. And that's what is happening now, again, when we have this AI chatbots like chatgpt and others. So they're heavily using vectors of words. And I use that same technique to create vectors for fragrances. And I think that the same applies as in NLP. If you have, you can create a single vector from a book, which has, I don't know, hundreds of thousands of words, but then this particular vector doesn't have some kind of semantic semantics that can be easily understandable. So that's probably the same for... the same could happen for fragrance vectors. So if you create a vector out of, I don't know, couple of hundreds of or maybe even more, thousands of perfumes, and then this one, this single one maybe is not... it doesn't point in some meaningful direction. So yeah, you could though you're having a lot of data, you are not getting anywhere basically. So it's better than in this if you unite all these data. So it's basically better to look at it separately. Yes, exactly. smaller, maybe in smaller subsets. Okay, yeah. Subset is a new word. Vector is a new word. mean like, our listeners gonna kill us probably by this point. Yeah, going back to the practical applications. we know that we could, like I'm an owner uh of a small brand. I have no one working there. So I could come to you and you could recommend me a perfume based on what I want to achieve. What else could you recommend? I could summarize the information from your existing perfumes and I could compare this information with the uh averages from the industry. For example, ah I could compare, I could see how your uh average usage of notes and accords compares to the industry from, let's say like uh last five... years or some other time span if you want. And I can compare your brand with other brands if you want and to see what are similarities and what are differences. This is more analytical and statistics based approach. I don't use a lot of, uh let's say, uh fancy data science stuff like vectors. here but I could do those comparisons so to analyze your brand and compare your brand with other brands to see how similar or different they are and also uh then you could see how are your current characteristics aligned with your goals and what you have to change or do differently if you need. if I have any goals. I could show you the trends, some trends in the industry also, in terms of used notes, accords, what do people like more, what do they don't like maybe and so on. I remember the infographics you put on LinkedIn about the words which are most like the words which are like everyone shared these infographics. So the words which are used the most in the fragrance industry, like not in the fragrance industry probably but in the perfume descriptions. in the names, in the titles. I'm the title. Perfumes. yeah, rose and ouds de minets, right? Yes, yes, exactly, exactly. Okay, Milan could show me trans and he showed me trans. And I'm a kind of person which when I see trans I start thinking like maybe I should do something the opposite. The thing is that like with Angel the perfume, know, Olivier Cresp was probably like if he had a tool at his time which showed him that this is not the trendy perfume and we are not. accepting it to the market, this perfume would have never answered the market. Because he did something new. And I think like it's my fault, but I just want to hear what you was thinking about when you've created this possibility. Is that ah for me, the value of knowing trends in any industry is about understanding what I could do a little bit differently while relying on them. trends so that it would be something super cool. um What do you think? What's going to be the most common application for your clients? Yeah, mean, those things which I can analyze could be used exactly as you mentioned. So ah you could see what is happening now and uh do the actions uh which you want to maybe to create something what can happen in the future. ah that something is not similar to the things which are happening now. I would not say to anybody, you have to do this like this and this, could just say what the current state is and the clients are free to do wherever they want. I don't consider myself as some kind of uh advisor. I'm only showing what is happening now and what are the facts from the data. And yeah, but you could, you could, mean, everyone has uh its own freedom to do whatever, whatever they want. And if you want to do something with this trendy now, you could do that. And if you want to create something opposite, you can also do that because that's actually how the trends are changed. once I hear that a trend is your friend. But some people add until it lasts. yeah, if you want to be ahead of the trends, then you definitely actually need to do something differently because the trends are cyclical and something what is not trendy at the moment can be trendy in the future. yeah, so but what I can do from the data, I could say, OK, this is the current situation. And that's what I see in my data. then I could, mean, again, it depends on the goals you have. Yeah, why did you collect these data in the first place, right? Because if you want to make, I don't know, a market analysis, then your data is just... Because I did it manually and it was pretty awkward. But I think that if I had your power at that time, my graphs would look differently. it's just so awkward what you do without the data. Yeah, mean, the data can help, definitely. I like when I have the data uh so I have some kind of proof that uh I can say, that's not something that I came up with randomly, but that's what I found in the data. Yes. I totally share it, because I know that lot of new brands, they go to a shop and they just smell some perfumes and say, it's probably the most common trends. However, it might be the most common trends of this particular perfume shop or of this country or... Technically you are like everyone as a human have very limited idea of objectability. Is it the word we could use? So what is true and what is not. And also we have a lot of mental tricks which sometimes give not very, I would say truthful result. So yeah, it's cool to have something which would support your idea of something or will reject or maybe you shouldn't have any idea before coming to your agency. ah Yeah, I mean, uh if you don't have ideas, I could show you what is happening now and maybe I could inspire you to have some. Now we know that we could uh find a perfumer, could find trends, ah we could find similar brands based on the data. Maybe some other services your clients could account for. If you want to produce fragrances, I cannot do that now, but if you are a fragrance producer with the data with exact aromas and formulas you used in your previous fragrance products, I could create custom AI models which could say, okay, if you want a fragrance with these notes, just the list of notes as they are in the app, uh Okay, then I could predict the exact amounts of ingredients to create something like this. And then the perfumers could uh start with that and then they could test those formulations further. So they don't have to start from scratch, but they could start from this point, from those predictions, and then they could refine them. But that's more for bigger fragrance producers, which have those data because I don't have it. Yeah, so I could create some custom models for their own specific use cases. Uh-huh. You uh do expect different or you do have different categories of clients. started, if I got it right, started from small houses or even like marketing agencies probably. Yes. To uh creme de la creme. So the houses which also, which have a lot of data. Yeah, because the data, I mean, if we are talking about AI, the more uh quality data you have, the better. So I think that there's a higher probability that the bigger fragrance producers have more data. mean, maybe that's not a rule, but this is my intuition. And yeah, so I could work with uh smaller niche perfume houses to show them the trends and maybe to recommend them. perfumers, as I said, or with big producers, which do have more data to create custom AI solutions for them. My intuition says that some of your AI models would be probably in demand at the big houses. People want to have these data and want to have them customized once they could afford it. So that's why I think it's very special about what you do. just a tool which you could ah develop to any direction. That's the beauty of it. mean, it depends on the data, depends on your goals. uh yeah, I think I could do the collaborations with both sides, let's say, with big ones and with smaller ones. because I believe that if it's the right tool, it could be used differently. Yeah, it's important that if you want something customized, then you need to have your own data. And if you don't have it, then I can show you what are the trends from my data. yeah, but everything is actually, the data is a key. So on both sides, we need data. So it depends who has it and how much. Yeah, a new gold. Mil, what are the best ways for the listeners and for the customers to oh find more information about your products and what you do? Well, now I'm uh the most active on LinkedIn and I'm trying to create weekly posts there about different uh data insights I found in my data. I don't have others data, then I do analysis and I post the results there. So I'm active there. uh Yeah, you can visit my website. I described some of those services, which we talked about here also there. And yeah, in the future I plan to go to some exhibitions and congresses. Right now I'm a little bit busy with some other things, but yeah, that's also the plan for the future. And yeah, so I do have a... the other social media as well, but I'm most active on LinkedIn now. oh share all the links definitely. Before we uh wrap up I wanted to ask you a few please questions. Those mean are just you should give very quick answers. So the first one. Favorite perfume scent? ah Something aquatic like the sea and citruses I see I think Do you think about, it's not on the blades, but do you think about the way aquatic vector look like at this point or not when you're saying? Yeah, it can be translated into vector. Yeah, but I actually don't think in terms of vectors when I'm thinking about sense. uh I'm like any other people. Yeah, I have visual associations in my head, not the numerical ones. Beautiful. Right, the second one. One word to describe your approach to data analytics. systematic It's not getting easier, but okay. The dream client company at the moment. Oh, that's a tough one. Let's say one uh Italian fashion designer. uh Wow, not bad. All the fashion designers in Italy are right now starting thinking, who is that one lucky one? uh Number four. What do you do to unwind after a busy day at Santa Aledis? Yeah, I go to the gym. Very good. Very appreciated. And number five, if you could analyze data for any other industry, like not for the fragments industry, and you have, it's just adding now to what you said, the beautifully collected data, which you could just like use for any model training. What would be the industry or maybe the type of application or... Maybe the data from stock trading. yeah, I mean, that would be more from. I would like to see if I can develop some useful models for stock predictions, for stock movements predictions, but I mean, that's from more, ah again, from more analytical ah approach. ah yeah, I'm not sure for uh nothing came to my mind now. But this one is very practical, I think. Yeah. Beautiful. Well, Milan, thank you so much. Thank you very much uh for the talk. ah It was very nice to talk to you and it passed very quickly. Yeah, it's always like that. Thank you. Yeah, thank you. Milan, you're so clever, I mean... you