Episode 30: How Tastry "Taught A Computer How To Taste" with CEO Katerina Axelsson

Tune in to this episode of AI at Work for a conversation with host, Rob May, and Katerina Axelsson, CEO and Founder of Tastry. At Tastry, they use innovative AI and flavor chemistry to generate singularized recommendations to shoppers for sensory-based products. Katerina shares how they "taught a computer how to taste", her unique journey from chemist to CEO, lessons she has learned building an AI company, and much more.

Subscribe to AI at Work on iTunesGoogle Play, or Spotify

Rob Circle Headshot

Rob May, CEO and Co-Founder,

Katernia Axelsson Headshot

Katerina Axelsson, CEO and Founder,

Episode Transcription   


Rob May: Hello, everybody. Welcome to the latest edition of AI at Work. I'm Rob May, the co-founder and CEO of Talla. I'm your host today. Today I have with me Katerina Axelsson who is the CEO of Tastry. And full disclosure, this is one of my angel investments.

I think you're going to find it really, really interesting because Tastry plays in an area of AI that not many people are looking at right now. Kat, welcome to the program. And I'll let you give a little bit about your background and how Tastry came to be. And tell everybody sort of like what Tastry does at its core.

Katerina Axelsson: Thanks for having me on, Rob. I'm really excited to be here. So first, Tastry is SaaS and insights company. And we're what you would call a multidisciplinary innovator in machine learning, analytical chemistry, and flavor chemistry. Basically, we created methods to predict how sensory based products, like wine, beverages, food, and fine fragrances would be perceived by consumers in general and down to the specific individual in particular.

At Tastry we liked to say that we taught a computer how to taste, because our core innovations are in our ability to break down the flavor matrices of these various products and relate them to consumers.

RM: So this is very different than the way sort of most recommendation systems work today, right? That's a big part of your insight.

KA: We do have a bit of a different approach. We created it because we had a harder problem to solve, I think. Yeah.

RM: You were a chemistry major.

KA: Yes.

RM: Out of Berkeley, right?

KA: Cal Poly. It's in the central coast of California. It's an engineering school.

RM: How did you go from being a chemistry major to being an AI person to this idea behind Tastry and realizing that traditional recommendation systems weren't going to work for this problem?

KA: It happened because I made some naive observations while working in the wine industry. Cal Poly is in wine country. We have 400 wineries in like a 20 mile radius. To pay for college I worked as a chemist in a custom crush facility where we would make wine.

I would notice that it's very common to have a 10,000 gallon tank of wine, sell half of it to one customer, and half of it to another customer. The same product would go under two different bottles, two different labels, and have different prices. Then, it would receive a different score from the same critics. I thought that there might be an opportunity to create a more objective, insightful system to connect consumers to products.

I had this hypothesis and I got permission from the owners to be a mad scientist in the lab until 2:00 or 3:00 AM for like a year. I did that and I developed a unique data set that I later discovered. I did some funky things and ended up creating unique, innovative methodologies for measuring those products. I read a lot and I knew that I wanted to learn about alternative methods to collaborative filtering algorithms, because the limitation was, I didn't have a lot of data.

I had a little bit of interesting unique data, but I didn't have a lot of users to derive insight and things like that.

RM: Let's pause and define for people listening. Collaborative filtering is the most common recommendation algorithm today, which says, let me build a vector, Kat likes these 20 books, Rob read these 20 books, 18 of them are the same, they'll probably like the other two that each of them read, right? It'll cross recommend them.

KA: Yeah. Exactly. That's a good point.

RM: You need a lot of data to do that successfully.

KA: And actually, it's not very helpful for sensory based products to do that. I mean, I can behave similarly to someone in the same demographic who buys the same things on Amazon, but we can go out to dinner and behave the same way-- and dis-similarly, and she could have loved that chardonnay, and I could hate it. There needs to be an extra layer of complexity for a sensory based product, so you needed a different data set as well.

I set up a meeting with this computer science PhD at Cal Poly just to kind of answer some of my questions. This half hour meeting turned into a four hour meeting. He brought two other PhDs into the room. They started arguing and drawing logic trees on the board.

Long story short, he's been our CTO for the past 2 and 1/2 years. We patented and validated the ability to track and predict consumer preferences for sensory based products.

RM: So, it works differently because it's really targeted towards sensory products. The example I always give people is, just because I know something about what you like, you both like steak, but not salmon, for example. That doesn't necessarily mean there's a correlation between something like, do you like chocolate and mint together?

Sensory flavors work a little bit differently. What interested me about Tastry when you got started was this idea that this is a really different powerful technology that has tons of applications. How have you gone through the process of figuring out where to apply it and where to start? You could do so many different things. Talk about some of the things you're thinking about and exploring and some of the things that you're very concretely focused on today.

KA: So, because the core innovation was predicting and identifying consumer preferences, there were a lot of of use cases and a lot of potential customers we could have brought this to. I started off by creating a recommender that would recommend wine to shoppers in the grocery store. I thought, once we have enough data and insight there, that we can move to manufacturers and producers and distributors. That part came a lot faster than we expected because the core competency was there.

We delivered a lot of value to a lot of different customers along the supply chain. That's because we broke down the flavor matrix of sensory based products in general. We didn't just train off a dataset of wine. It was about the chemistry that matters to humans.

RM: Does this approach also get better with data the more that you have?

KA: Yeah. I think the data scientists will always say they would've liked more data. We do have a lot of data on the chemistry side. A lot of complex data. It's a very good problem for machine learning to solve because of all the inter-dependencies. There are thousands of compounds in a single product. We're big on data on that front.

RM: Right. When you look at it in the long term, how much are you thinking about becoming a company where you're going to sort of license technology to third parties to help them determine flavor profiles versus produce your own flavors or your own products or be a different kind of recommendation system? How are you evaluating those opportunities?

KA: We've been to a lot of trade shows and we actually found that very large flavor and fragrance companies and distributors came to us and they told us what use cases they found compelling to use our technology for. We've been engaging with them, we've evaluated efficacy with them. We have identified some very interesting use cases around product development. We ran a test and we found that we could actually predict the aggregate consumer score for a wine before it hits the market with north of 90% accuracy.

We can determine and isolate the variables that are responsible for that score. That's very interesting to manufacturers because a tenth of a point rating from a consumer score vastly increases the retail value of the product. If they can modify it before it hits the market, they consider that a new approach to the way they currently do things.

RM: Very interesting. Talk a little bit about some of the concepts that you have explained to me about the business that listeners might not be familiar with. Like, what is a flavor matrix? What is singularization?

KA: I'll start with singularization. We felt like we had to create the term singularization because personalization and hyper-personalization didn't quite capture what we were doing. I felt like when I was talking to retailers about their personalization program, they were like saying, oh, so you bend people into big bold red wine drinkers or sweet white wine drinkers. And I said, no, it's much more complex than that. We can actually identify how an individual will score a product to a tenth of a point.

That just seemed like a really novel thing to them. So, now I say singularize because we can make predictions at the individual level and the group of consumers level and a local and regional level. Then the flavor matrix is basically the relation of the different compounds in those products. The example I like to give is a benzaldehyde. That's one of the many compounds responsible for the flavor of cherry in a red wine.

But, not every red wine is described as having that flavor. That's because of the presence or absence of hundreds of different compounds that are allowing it to be masked or expressed. We basically use machine learning to break those relationships down and figure out what matters to consumers. Then connect it to a consumer profile preferences.

RM: One of the pieces of advice that I give to executives sometimes who are thinking about where to start applying machine learning in their companies, I'll give the example, I'll say, is there's something where you could in theory untangle the rules of how something works, but there are so many rules that it would be so complicated that it would take way too long to be worth your effort, or something like that. That's a good problem, often, if you have the right data set up for machine learning. Do you feel like Tastry sort of falls into that bucket?

KA: Yeah. Absolutely. And, I would also say humans don't do this. We're not really replacing anything anyone is really doing. Like, a sommelier doesn't do this. They have a very different job. We're doing something humans can never do, which is identify these inter-dependencies and complexities in a product.

RM: So, this is one of those use cases of AI where you're not taking jobs, you're not putting anybody out of work, but in fact, you're going to allow companies that are food companies and beverage companies, companies that intentionally had to mass market their products effectively to get any kind of scale. Now you're going to allow them possibly some day to like singularize them, right? Or mass personalize them.

KA: Yeah. Exactly. And in that way it's also disruptive because we can change the way, for example, a product is made during the development cycle. So, whereas before you'd have to make a product, wait and see how it performs in the market. Now, we can simulate how it would perform and make modifications before it hits the market as well. No one was doing that before.

RM: Very, very interesting. Let's get back to the fact that you were a chemist and not an AI person. Obviously now, you've been doing this for a couple years. So, you've had to come up to speed. You're the CEO of an AI company. What's your advice for people that are trying to get started?

How do you think about it? Are there any key lessons you've learned, or frameworks or sort of key ideas that you used that you think people should look into first? How do people get started in AI when they're not super technical or they're not computer scientists?

KA: I would say part of it was I was fortunate to be surrounded by a lot of very, very smart people, a lot smarter than I am, who are very educated in this space. A part of it was we were doing something. I wasn't sitting and learning. I was figuring things out as we were building the company. That really helped.

Your content is obviously very helpful. I would also say I think part of the benefit for me at least was that I wasn't a PhD in machine learning or something like that. I think the lack of that kind of rigid thinking, helped new ideas come up and new approaches come up.

RM: Yeah. Definitely. I mean, we've seen a similar thing, right? Definitely, in some cases, PhDs can make good entrepreneurs. More often than not they don't, because I think the kinds of things that draw you to academia, it's a very different type of problem than the problems you have building a company. Company building is very pragmatic and you have to like the nitty gritty apply problems.

It's not about, how do I come up with this genius model to solve this problem? A lot of it is like, how do I get this model to run in the memory that I have available? Stuff like that. Is there anything that you guys talk about that are maybe AI problems that you're not working on that you wish somebody else would solve to make your life easier?

KA: If there was a more data driven approach in general with some of the industries that we're kind of trying to work with.

RM: I mean, is the consumer goods industry for flavor related stuff or sensory related stuff, is it primarily driven by a handful of people who are just the experts on this? Is that how it works?

KA: I would say the biggest challenge is the status quo. So, what we're really competing against I would say is, like an expert in the industry who's on a retainer to come into a winery and taste a product and then give recommendations as to how to make modifications to release it into the market. So, we're basically doing that, but with a data driven approach and with a guarantee.

RM: Sort of like AI Moneyball for consumer food manufacturers.

KA: Yeah. And there's a typical 80% failure rate right now for a new product that you launch into the market. I think a lot of these companies are very interested in mitigating that risk in any way they can.

RM: That makes a lot of sense. I mean, we've seen-- I've seen across the board and the people that we've interviewed on the program here, there's tons and tons of inertia to overcome to adopt AI. I think what's going to start to happen is you are going to see this sort of Moneyball thing play out where somebody who is maybe like not the market leader or whatever is going to be like, OK, well, I'll take the chance because maybe it gives me a chance to be number one.

As that works out, which I think it will for most of the people that do that, then other people will be forced to follow along. I mean, all the teams do the Moneyball thing now in baseball, right? You're always looking for the next edge. It's very, very similar in business, I think. Are there other applications of AI that you see out there that you personally used that just really, really get you excited? Is everything tangential to your industry that's also sort of like food and beverage related that's AI driven.

KA: Not really. I wish I had a better answer for this. I wish I could say, oh, I have competitors and this is how we're different. But, I haven't really seen anything there. I would say, I could see other applications for this type of approach that would be really exciting in other fields, like bio-informatics or something like that. I would say I'm just really focused on our method and what we're doing. Every day I'm learning more and more on how it's unique. 

RM: As the CEO of an AI company, what have you learned about like hiring machine learning talent, data scientists? Has that been easy? Has that been hard? Do you have any lessons for the people listening?

KA: I would say that I hire for personality first. I think you have to just kind of be a certain type person to solve a new type of problem that you've never seen before, never learned about in school before. There is a shortage of data scientists and machine learning talents. If you find the right person and you can train them, I think that we've been very fortunate with that.

RM: Definitely. It's early 2019, where's Tastry going? Do you have a lot of goals that are related to the technology? Are most of your goals more on the commercial side?

KA: We've developed quite a pipeline. 2019 is all about executing on that pipeline. We've educated customers and different people in the industry on what we can do. I'm really excited to get all that going. My dream for 2019 would be to, you know how IBM beat Kasparov at chess, right? I would like for Tastry to kind of identify various wines alongside a world class sommelier.

RM: Interesting. That would be a great marketing tactic, to get something like that setup.

KA: Maybe on the Food Network or something.

RM: Last question before we go into some of our more general AI questions. How do you know, so you guys have this theory that this is better, consumers concerns, but like, what do you do to prove that this is effective and it really works?

KA: Our favorite thing to do is run an efficacy test. As soon as I get a meeting, one of the first things I say is, typically, we're underestimated in what we can do. So I just say, look, that doesn't matter. Let me just prove it to you.

Usually I have people send me an unidentified product or a product that they have a lot of data for that we wouldn't have access to, and we would identify how that product was either rated or what variables in that product were modified or something like that. That usually does the job. So, demonstrate. I would say less talk and more demonstration has worked.

RM: I think a lot of AI companies of different types are running into the same thing. It's the reason that a lot of so many companies are doing pilots trying to figure this stuff out. But, yeah. Well, cool. It sounds like you guys are onto some good stuff.

One of the questions that last year we ended all the podcasts with this Mark Zuckerberg versus Elon Musk question, but that's old news now. So now, we have the Gary Marcus versus the rest of the world question, which is, you have Gary Marcus out there sort of saying like, hey, deep learning is only going to get us so far. We need new techniques to really have breakthroughs. And, you have a lot of the world saying, no, deep learning's got a long way to go.

Some people think deep learning is going to get us much further. Some people think not that much further. Do you have an opinion on any of that? Do you have an opinion on the future of AI and where breakthroughs need to occur, or who might be right in that debate?

KA: I'm going to give you some insight maybe as a chemist. I would say, I think there will be more AI companies that have a multidisciplinary approach. And if you have that with deep learning you add, for example, in our case, chemistry with deep learning, you might be able to gather some interesting data sets that would be able to move machine learning AI forward.

RM: Well, you see, of almost every serious production level AI system is actually a ensemble of various methods. You see, I mean, we even do this even at Talla, right? We have sort of this machine learning pipeline that you put things through. First you're trying to figure out a model that just like, what is somebody trying to do?

Are they trying to ask a question? Are they trying to execute an action? Are they trying to do a search? Now that we know what they're trying to do, or at least we think they're trying to do, passing to a new model that does that thing. So yeah, very cool.

Well, Katerina Axelsson, thanks for being on the program today. If people want to learn more about Tastry, what's the best URL to use?

KA: It's just Tastry.com. It's like pastry, but with a T.

RM: And if they go there, can they still try the wine? Do you still have that there?

KA: Yeah. It's like a technology demonstrator right now, but you could create a profile, get a wine recommendation, save it.

RM: I will just say, as we're wrapping up, that part of the reason I invested in here is when I first sat down with Kat and she had me fill this out, the number two wine that the algorithm recommended is the wine that I am most likely to order if I'm out at a restaurant. I was pretty impressed because you couldn't mine that data about me. I don't talk about it on social media or anywhere else.

It was pretty interesting. Kat, thanks for being on the program. For those of you listening, if you have guests you'd like us to see, or questions you'd like us to ask, please email those to podcast@talla.com. We will see you next week.

Subscribe to AI at Work on iTunes, Google Play or Spotify and share with your network! If you have feedback or questions, we'd love to hear from you at podcast@talla.com or tweet at us @talllainc.