#3 Carlos Fenollosa – Founder and CEO at Optimus Price on running an AI startup in Barcelona during the COVID-19 lockdown

A researcher and data analyst at heart, Carlos Fennollosa spent 7 years researching predictive models for DNA and protein modeling at the Barcelona Supercomputing Center. Carlos founded three startups, the first two failed and the last one is growing strong: https://optimusprice.ai

#3 Carlos Fenollosa – Founder and CEO at Optimus Price on running an AI startup in Barcelona during the COVID-19 lockdown Get Your AI On!

Ciprian Borodescu: I’m here with Carlos Fenollosa, Co-founder and CEO at Optimus Price. I’m super excited, and it’s an honor to have you on this podcast. Thank you so much for being here.

Carlos Fenollosa: Thanks a lot, Ciprian. It’s a pleasure to be here, in your podcast, as one of the first participants in this new podcast.

Ciprian Borodescu: So I suggest we dive right in. Why don’t you tell us a little bit about yourself? And I have to say – you know this already – I love the name Optimus Price, and maybe if you want to share the story of how that came to be?

Carlos Fenollosa: Yeah, of course. So, Optimus Price is my third project/ startup. My story is quite straightforward. So, I’m a computer engineer, I did a Master’s in artificial intelligence at the Barcelona Supercomputing Center, which we are very lucky to have. It’s one of the best supercomputing centers, both in terms of the research that it’s done and the machine we have – the Marino’s room. I started a research career there. I did two years of my Ph.D. And I wanted to do something different. I kind of always been curious about the startup ecosystem, you know, watching movies about Steve Jobs, Bill Gates, and the Silicon Valley. And somebody said, “Can I create a startup that is about research?” So research, but applied to some business problems.

Carlos Fenollosa: And I must say, Optimus Price is the third project. Before that, I did a project that was more related to biotech and technology, which is what I was doing my Ph.D. on. It didn’t work out, we had to close. Then I started a second project, which was a chatbot, a conversational bot. And I did what they call the engineer’s mistake in a startup, which is to close the door and stay in your room for a year, develop some very cool technology, and then go to the market and try to sell it. That is what kills most startups – and keep in mind, I went outside, nobody wanted to buy what I had developed. But, at the end, you know, the chatbot was used for some projects, not very interesting, but at the end, I found somebody who wanted to try part of the technology, but not the bot. So, I thought, “The bot is the cool thing.” And he said, “No. Look, the cool thing is that you told me that this bot can do analytics over sales”, which was one of the features of the bot. You could connect it to analytics and it would give you a sales forecast. We tried the sales forecast within that data, the accuracy was very high, and then this potential customer said, “Look, I will buy your product, but not the bot. I want a normal web page with a normal plot and a normal table and normal buttons so that I can analyze my sales data better.” And this is how Optimus Price has started.

Carlos Fenollosa: And the name, I mean, it’s obviously a pun on Optimus Prime. Our first logo was actually a vector image that I drew myself of Optimus Prime, but at some point, an investor told me, “Carlos, you should change this because you will get sued.” And we decided to put this kind of funny name to have an impact – a name that tells you something about the company. So, the founders, I have to say, we had a strong discussion, I was the only one promoting this name, my colleagues wanted a more normal name. But at the end, I said, “Look, if we’re going to crash, we’re not going to crash because of the name. I mean, people are not going to discard our proposal because the company name is Optimus Price. I think quite the opposite.” And it was a gamble after all, because you never know how people react to that, but it worked really well for us. So, now people remember Optimus Price, they associate it to price optimization, which is kind of the point, and it’s almost a conversation starter.

Ciprian Borodescu: Yeah, absolutely. Thank you for sharing that story. This is awesome. People understand immediately what you, guys, do. Cool. What are some of the common use cases where Optimus Price as a platform can add value?

Carlos Fenollosa: Of course, as optimization for retail, which is important. We have not entered other fields like, you can optimize prices for travel or insurance, for utilities, for the stock market. We thought to apply our models in those industries, but we decided not to. So, we are focused just on retail. And like I was saying, retail is a very old business, there’s not a lot of room for innovation in processes. Actually, probably the biggest win of the 21st century is the streamlining of the supply chain, but not the prices. And almost nobody is doing this with data. We have done a lot of research ourselves, we have looked at papers published in the industry, books, and even the most classical books, they use rows of three or just additions and subtractions and multiplications – there isn’t a lot of science into that. So, when we started delivering this software that optimizes prices, we discovered a new thing; in retail, there are a lot of processes where you are losing money. And we saw that optimizing prices was one thing we could do, but our models could be also applied to other processes – the same models with very little modifications.

Carlos Fenollosa: So, we entered into a second use case besides pricing, which is to optimize stocks and the supply chain. So, on one side, you can set prices by calculating price elasticity, so that you are going to increase your margins, which is nice. But on the other hand, using those same models, you can detect which is the optimal amount of stock you should have in every store, you can have dynamic safety stock – safety stock is a concept that means the minimum amount of products you have in every store so that if some customers come to purchase an item, you don’t say, “Sorry, we are out of this item.” So, you have a reserve, which is typically some number. If you are very technological, maybe you use the average of the sales of the past three months, or maybe you have a fixed number, and say, “Of every item I have in the store, in my back office, I have five of them, just in case.” So with Optimus Price, you can use a dynamic number. I mean, if you are sure that you will not sell more than five in a month, why have more than five? Or if you see that there is a peak in demand, you can move more to the storage so that you don’t run out of stock.

Carlos Fenollosa: So, the models are very versatile and, as usual, with a technological company, the problem is finding use cases. So, a common use case is to see which competitors you should follow when they lower prices and which you should not follow. So, right now, there is this race to the bottom, especially in eCommerce, where you can see this behavior online: somebody has a product, which sells at 10. And then one competitor puts it at nine, and then another automatically puts it at 8,99, and another puts it at 8,95 – and a lower price is not necessary. So what we did with our models is, which is the competitor really that you should look at – not all of them, but the most important for you – and if you can afford to be more expensive. Actually, this morning, I was in a conversation with a client who told me that this was the feature that they liked the most about Optimus Price, to ask this question, “Can I raise my prices and still sell and make money?” And not for all products and not in all cases, but sometimes you can. So, while the system tells the retailer to lower the price for some products, it also tells them to raise the price for some other products, and they make money for both cases. And sometimes you don’t have complete information or the data is not clear or you are selling your own brand, so you cannot compare. If you are, for example, a fashion designer, you cannot compare your white shirt with the white shirt of the competition because the materials may be different, the patterns are different. So, it’s not like it’s a one-on-one comparison. So, essentially, the models here work much better than models that are based on competitors.

Carlos Fenollosa: And finally, the last thing that we have done very recently is to link both worlds: the demand and the price elasticity so that whenever a product is in high season or is in high demand, we check if you can raise prices to make more money or, on the other hand, you have to lower the prices. So, this dynamic pricing system takes into account a lot of variables that pricing managers, which are the users of our tool, already had in their mind but they cannot operate on them because there are lots of rows, lots of columns; and I know we will talk later about data because this is the core problem – there is just too much data to do this by hand, even with Excel, so you really need AI, ideally, or at least some smart algorithms that can help you balance demand with your prices.

Ciprian Borodescu: Okay, okay. This sounds really interesting and I’m wondering what is the lifecycle of a project involving Optimus Price? Where do you start? What are the usual steps in the timelines for onboarding, let’s say a new client until they get to production?

Carlos Fenollosa: Well, getting the data is the most difficult part – working with the data. And we have seen this for different types of clients. I will try to make this generic because I don’t want to talk too much about the specific case of Optimus Price. I think it is a more general problem for all the companies that work with data.

Ciprian Borodescu: Yeah, absolutely. We also have that. Yeah.

Carlos Fenollosa: Yeah. And I have spoken to our own competitors, the Optimus Price’s competitors, to friends that have data companies in other industries – it’s always the same. First of all, let’s assume that the client has signed the contract, we are very happy, we open a bottle of champagne, and we have a toast because we brought a client. So, now we tell them, “Okay, you have to give us your data because otherwise, we cannot work with that.” And the client has promised you that all their data is okay, they prepare a data feed, and then you have to analyze it. Because even if they swear to God that the data is perfect, you’re going to find incorrect formats, missing data, missing periods of time, variables that are improperly placed, strings where they shouldn’t be, or floats that sometimes have commas and sometimes have periods.

Ciprian Borodescu: Absoultely. Yeah, 99% of the time.

Carlos Fenollosa: Yeah, I mean, how can this happen? But it happens. So, you have to normalize a lot of the variables, depending on the impact of everyone. So, feature selection and weight selection are usual in machine learning. And then when we have said that, we can train the models with the historical data so that they can do forecasting and prediction of future data based on historical labor. And then, you put the account in production, but we are going to find some edge cases because there are always edge cases. And I would like you, the client, who is the business expert and has the knowledge of the domain to let me know whenever you find some strange prediction. Different business cases always have different behaviors and strange things happening with clients. Because in the end, what we are trying to predict is consumer behavior – which is totally irrational, as you know.

Ciprian Borodescu: Absolutely.

Carlos Fenollosa: The last part is fixing those edge cases, which are not a lot of them, but they cause trouble and to the eyes of the client, they don’t understand what’s going on behind, so they typically complain. In 90% of the cases, you discover that there is a problem with the data, but 10% is bad behavior of the models.

Ciprian Borodescu: Yeah. And this is one of the things we’re also getting asked by merchants: are you aggregating my data with other merchants to build your models? And if not, how big of a data set do you actually need from me to train your models? What’s the situation with you, guys? How do you deal with these types of questions?

Carlos Fenollosa: Yeah, we get these questions a lot. People are worried that competitors might benefit from their data. But no, the models are independent. When we say we have a generic one is that we have regression functions and normalization functions, which are, I would say more important than the actual predictor, that have been built with the business knowledge and the domain knowledge from many clients. So, right now, with the Coronavirus situation, we are trying to understand how can we build models – and if that is even possible – that can react to the changes in demand and the changes in consumer preferences, you know, just with the flip of a switch. Can we do something like that? Like, have a global pandemic variable that is used to adjust these kinds of models? So, that’s what we are trying to do. So, we use data from the clients, of course, always anonymized, and we never get personal data, so we don’t have data from the final consumer or anything; we only know about products and sales.

Carlos Fenollosa: But we try to learn from that and to do research because one of the topics I was mentioning before, there is not much research done in the retail industry, and for sure none has been done with data. We want to see if some kind of patterns emerge. And just to conclude this point, one of the patterns we have seen is the weather. So, the weather changes a lot the consumer’s behavior, and especially they change from summer to winter, and in between periods if it rains or not. So some of our models – I mean, not all of them, but for some clients – we include the weather variable, so that for clothes and other kinds of products, like sodas, refreshments, alcohol, ice cream, products that are very typical for the summer, the system learns when summer starts; and summer not in the geological sense or in the global weather sense. But rather, when people start to feel that it’s getting hotter, therefore I drink more beer because it’s socially acceptable or for whatever reason. That is very interesting. And definitely is out of work and the day-to-day typically overruns you and you can’t afford other resources to do that. But data aggregation, at our level, is really interesting to do research and to understand things that even retailers or clients don’t know how they work.

Ciprian Borodescu: And do you need a certain dataset size from your customers? What kind of companies are, let’s say, best suited to benefit from Optimus Price?

Carlos Fenollosa: We designed our product since the very beginning so that it could work with very small amounts of data, with very small datasets. Of course, you cannot do miracles, you cannot do forecasting if a product is sold once a year; it is impossible. But what other things can you do? So we invested a lot in connecting to other data sources and trying to be very creative to enrich those models with metadata or what we call fake elasticity. So, elasticity is how demand changes depending on the price, but really, the demand if you only evaluate it with sales, the final sale is a very precious data point, and you don’t get a lot of them. But you have other intermediate data points that can help you detect the behavior of the consumer – for example, website visits, adds to cart, interaction with the website, searches in Google or in other search engines, interest in Facebook; if you run ad campaigns, the impact of the ads. So because we wanted our tool to be useful even with very small datasets, what we need is to get more data – lower quality data, of course – but more data from other different datasets so that for a small store is very difficult to optimize prices and calculate demand for the long tail of products and even for top sellers in some cases. So, what we do is cross-reference it with Google Analytics information, with traffic, with ad campaigns, with a general interest for products, weather – I mean, any data point we can use, we use it.

Ciprian Borodescu: So instead of actually having the sales themselves, you have indirect data that points to eventual sales. Right?

Carlos Fenollosa: Exactly. And that data has a bias because not every visit to the website is a sale, of course, but it’s better than having no data.

Ciprian Borodescu: Absolutely.

Carlos Fenollosa: And it works really well because we are trying to model, again, consumer behavior, which again is irrational, but the hints are very strong. And the conversion rate is a great indicator of this. If a product has a very bad conversion rate, and I could say worse than the average conversion rate of the category, you have a problem with the price. Because sometimes, the product doesn’t sell because it’s just not popular or people don’t need it that much – they only need to buy it once every year – or because the category is not popular because it’s old and outdated. But when you compare this product to a product in a similar category, if the behavior of that specific product is clearly inferior to the rest is because you have a problem in price. It’s not a problem in product visibility, in marketing positioning, in the product itself. No, you’re just not selling it because it’s the wrong price. We can cross-reference it with what the competitors are doing, which is a very interesting data point. If the competitors are selling it at a lower price and you have it more expensive, and you don’t manage to get sales, I mean, lower the price, because otherwise, somebody else will get the sales. So, yeah, expanding your dataset and creating information is a very good way to be able to work with very small datasets – very small, high-quality datasets. So, you increase the number of data points, and you lower the quality, and then the trick is how you weigh these new variables, and how you balance that so that you don’t introduce other biases.

Ciprian Borodescu: Yeah, I love the approach that you guys have. I mean, you know, we’ve been talking even in the past one or two years, when we went through the European Data Incubator, and one of the things that we do not do is external data sources, which I think you guys do in a marvelous way.

Carlos Fenollosa: And I have to say, we did not discover that by ourselves, alone. It was through many conversations with pricing managers. And our product initially was just based on price elasticity. So, of course, it didn’t work for small datasets. But we sat down with pricing managers of small companies, and we asked them, “Look, I mean, with your data, we cannot do magic. What should we do?” And they said, well, whenever this happens for products that don’t sell very well, what they do is to look at what the competitors are doing. Okay, so that’s interesting. What else do you do? Well, I look at how traffic is going. Okay, that’s also a good data point. What else do you do? And it’s the client who’s telling us what to do. And, at the end of this conversation, then, the client realizes, “Wow, there is a lot that I could do if I had the ability to check all these variables for the product.” But at the end of the day, the amount of work hours is limited, so I can only check this for my 10 or 20 top sellers. But if I have a catalog of 20,000 products, it means that the rest, the 19,000 something, nobody’s looking at them. And they could benefit from some kind of algorithm that does this exact process that a person would do, but automatically, and this is the trick of the AI with reinforced learning because sometimes your assumptions are not correct. So, whenever you build a model, if the model is not working, it quickly learns that the performance of the product is dropping, and then it goes back to a strategy that was known to work in the past. So, it’s not like you flip a switch and then the AI does strange things. You can see what’s going on and it self-corrects whenever it makes a mistake – which sometimes it does.

Ciprian Borodescu: Okay. What have you seen in general – and I’m talking about AI and machine learning – do you see retailers being more willing to embrace it nowadays, after the Coronavirus crisis? Where do you see the market heading from this point of view? And kind of what are the challenges because I understand that from a technical point of view – and this is something that we see as well – having access to the dataset is the first challenge, the first barrier. But then, what are the business challenges that you’ve seen, especially in this space?

Carlos Fenollosa: Seeing how now everybody on Twitter with an Excel spreadsheet can do relation models that predict infections and deaths better than every government in the world, I mean, maybe the general public is going to start believing scientists and data scientists more than before. AI is one of the pillars of data and decisions. It’s not the only one, but it’s one of them. So, it’s going to advance for sure. And in the next 10, 20 years, whenever tools are more democratized, so what everybody is doing now in Excel, they will do it in some kind of tool that probably doesn’t exist yet, that is like Excel for AI, or, let’s say, Excel for machine learning because AI is a wide field.

Ciprian Borodescu: I like that, Excel for machine learning. That’s awesome.

Carlos Fenollosa: Whoever creates that is going to be the next Microsoft. But I’m not sure how we can start to tackle that challenge.

Ciprian Borodescu: That’s very interesting. So, from our conversations with different companies, also in the eCommerce space, whenever we talk with a new company, then we have to convince the leadership team – whatever that is, like the marketing department, the CMO, or the CTO – that using AI is a good strategy for them. If we cannot advance from this gap, then there’s no deal, the company is not going to invest in innovation. Then if by some miracle, we come across a company that has great leadership and they understand the value of AI, and they also have the data, then there’s also the trust gap. And here the trust is characterized by, you know, is AI really adding me the right value? Is AI really trustworthy? And questions like that. And, of course, the next step is to hire talent. And you have in-house data scientists, machine learning engineers, and so on. So, that is the talent gap. And only far away in the future is the ethics gap. I have yet to come across companies that are questioning their methods with regard to AI. But yeah, this is how I see the AI maturity model being divided into these gaps that we identified.

Carlos Fenollosa: 100%, nothing to add. I think that you are right on the spot.

Carlos Fenollosa: Okay. So you guys, I know that you raised a round at the end of last year. Maybe we can end this podcast by telling us a little bit about your current struggles that you and your team are going through these days. And, of course, what’s the plan for the second half of this year?

Carlos Fenollosa: Our main struggle is that we’re working remotely. I have to say it’s not bad, but you really miss the amazing tool which is the office – having an office space.

Ciprian Borodescu: That’s a good one.

Carlos Fenollosa: For me, the office is just another tool. But it’s a very useful one. It’s not written in stone that you always have to be at the office. It’s not the best tool for all situations, but it’s a very good tool and a very efficient tool for most situations. That is why people work in offices because, if you don’t have the office you don’t have the water cooler conversations – and the personal contact and personal feeling are very important – now, everything requires more communication. Being remote means that you have to make a conscious effort to communicate with other people purposefully. Most meetings can be replaced by a phone call and most phone calls can be replaced by an email. But it’s true that for very important meetings to make very hard decisions, not being able to make it face to face, even with the best video conference software, I mean, there is no replacement. So, remote is difficult. And you’ve mentioned before hiring developers is very difficult. So, Barcelona now is one of the hottest startup ecosystems I would say in the world, and there is enough talent – I don’t want to say that there is not enough talent – but there is also more demand.

Ciprian Borodescu: One last thing I just remembered right now, what I saw on LinkedIn when you shared what you did with the interviews with the candidates? Can you tell the story?

Carlos Fenollosa: Yeah. This happened to me by chance, really. The recruiting process of a company is like the shopping window of a company. So my philosophy is that the recruiting process has to be as honest and transparent as possible, so that the candidate really knows what they’re getting into, and they can make a good decision if they want to work with that company or not. And also, you have to treat people as much as you would like to be treated. And the recruiting industry is this big irony that they call human resources, and nobody treats them like humans, they treat them like numbers.

Ciprian Borodescu: Exactly.

Carlos Fenollosa: So you send a curriculum, and nobody gets back at you, and you have a phone call, and you don’t have news later. So it’s really a big mess. So, because it’s difficult to attract talent, for a startup, you have to give some intangibles. And one of the intangibles you can give is to have a good recruiting process, to be nice to people, I mean, to be a good person, to show them where they are in the process, to provide them feedback. Nobody does that. But I do that because I think it’s the right thing to do. I mean, there is no other explanation. I just think that is how you should treat people.

Carlos Fenollosa: And on some Friday afternoon, I was very tired of replying to candidates – so I had spent the whole Friday looking for a position. And I have spent the full eight hours saving the curriculums and preparing custom replies to the discarded candidates. So of course, if a candidate progresses in the process, you contact them, by definition. But to discard it, sometimes you don’t send them anything, but I thought, let’s try to sweeten up the bad news. Because why not? I mean, I spent eight hours looking at curriculums, ranking them, looking at skills, everything. I can take a couple more minutes per candidate to write a custom email. And that’s what I did. I send about 80 or 90 emails. And this exploded, it got very viral on Twitter, I got calls from magazines, from radio stations to talk about this.

Ciprian Borodescu: Seriously? Really?

Carlos Fenollosa: I guess it connected to people. You never know when these things get vital.

Ciprian Borodescu: Absolutely. Excellent. But in all fairness, when I saw that, I immediately recognized that this is something I’m also passionate about. I recognized leadership. And I think this was an amazing thing to do, and yeah, keep doing that whenever you can, with respect to the people that apply to Optimus Price. And I think this is something that other CEOs and founders can also learn and apply to their own startup. I have to say, Carlos, it’s always a pleasure to talk to you. Thank you so much for being here, thank you for all your wisdom. Can you share how people can reach out to you for ideas and comments?

Carlos Fenollosa: Sure. Our website is optimusprice.ai. Besides my company, I’m very active in social networks. I have a blog, I tweet. So, you can look for my name, Carlos Fenollosa, or C Fenollosa, and you can look me up on Twitter or send me an email.