Better Data Science and AI Technologies for Better Vine and Wine?

Introduction to Wine

0:00Hi, everyone, and welcome to HDSR podcast. I'm Xiao Liman, the founding editing chief of Harvard Data Science Review. This episode is full about wine, and it's a part of our ongoing coverage of the Wine to Mine conference, which we started about a year ago. It's a conference about the use of AI and data for the wine industry. Before we dive in, I want to give a big congratulations

0:31to my usual co-host, Liberty Vitter. Liberty has a new roommate after waiting for nine months. She's taking some well-deserved time off to be with her family, and we were sending her lots of love, and also now Liberty is waiting for the moment to drink again. In the meantime, I'm thrilled to

Guest Co-Host McGill Paredes

0:55be joined by our guest co-host, McGill Paredes. McGill is a partner at Kearney, where he helps companies develop AI strategies and create real business value. He is also a wine lover and a co-editor for the column on active industry learning for Harvard Data Science Review. We're excited to be joined by two amazing guests. First is Katerina Axelsson, founder and CEO of Tastree. Her company literally taught a computer how to taste. Then we have Kia Bonilla, co-founder

1:31and CEO of Scout, which use camera technologies like your phone or webcam to monitor vine health, track performance, and predict yields. He's also the CEO and co-founder of Neotemper. And I can personally vouch their wines are really fantastic. We'll hear from both of them about how their companies are helping the wine industry to improve wine quality, boost production, and connect people with wines they will love. Let's get into it. Well, first of all, thank you all for

Kia Bonilla Interview

2:05joining this podcast. I'm going to start with Kia. Kia, first I want to thank you again for being a Tino for the Wine to Mine 2025. It was a fascinating talk. I'm also deeply grateful to you for hosting me at your winery. Enjoy your wine. But I'd like to start with you to ask a little bit about your adventure recently, what your company does, and how do you use data and AI. Just give a broad introduction to the audience about what you do. Thank you so much, Shali. So my journey really has

2:38three parts. The first part, I spent 25 years in my career really helping modernize and digitize a lot of different industries and bringing operational efficiency through the use of data. I was one of the early people at Tivoli Systems, Marimba, BMC Software, and then Splunk, my last journey. And what I realized is once you put powerful data in front of experts in their field, they know what to do with the data. My wife and I also love wine, and we bought a vineyard and started our wine journey

3:13about 14 years ago. And as we got into it, one of the first things I wanted to do was imagine the vineyard of the future. You know, how do I get the right set of data for our winemaking team and our viticulture team so that they could do the same? And what I found was a big gap. I took a course at UC Davis, my alumni, called the Wine Exec Program, met my co-founder, Mason Earls, who taught a course called the Vineyard of the Future. We started chatting about what's the best sensor in the vineyard, and he said, Kia, the answer is easy. It's the farmer's eye, because the farmer sees problems,

3:49remembers what last year was, understands what's different. The only problem is that doesn't scale. And that really became the inspiration for Scout. So what we have built is a solution that uses regular cameras, you know, much of which you can find in, you know, webcams, and, you know, it's incredible how cheap and powerful they've become. We attach that to regular smartphones, so no proprietary hardware, no expensive hardware for farmers. And that's important because we think

4:22this solution and this solution can expand to all vineyards and all farms, quite frankly, around the globe. And then the power of AI. So AI becomes the computational system to marry computer vision with the location of the vines. We take 20 photos of every single plant. We use the 20 photos to effectively create not only a picture, but also a medical record of each and every vine and capture

4:54the vital statistics, whether those are canopy and vigor growth, how many clusters, how many shoots. It's like an x-ray, basically going very deep using computer vision and labels with the power of AI to count things, measure things, and build a database, basically, of what you have. Kia, incredible work you're doing. Thank you so much for sharing. We also have Katerina. Katerina,

Katerina Axelsson Interview

5:19there was this really interesting CNN article that the title was, The Tech Startup That Taught a Computer to Taste Wine. We'd love to learn a little more about yourself, your journey, and tell us a little about Tastry and the problem you're solving. Thanks, Miguel. So before Tastry existed, I was a chemist, a very young in my career chemist, still going through college and paying my way through college by working in the wine industry. And I was always awed in that I was obsessively reading research

5:54papers. And because of the work I was doing behind the scenes in the wine industry and my exposure to it, and the fact that Cal Poly, where I went, was offering winemaking classes, I got interested in sensory science. And as I was reading these papers on all the focus groups that we do, all the ways we understand consumers and how they perceive chemistry, I started to feel unsatisfied. And like, there was a lot of gaps in understanding how consumers actually perceive, you know, flavors and fragrances and things

6:30like that. And so long story short, I had this thesis that the reason that we cannot have a consensus between predicting how much consumers would like something or what words they use to describe the experience is because machines don't look at chemistry the way humans do. We run panels, but we as humans are experiencing all these chemistries in different concentrations at one time as this complex chemical

7:03soup. And in wine, it's even more so because it's an agricultural product with a lot of nuances involved. And so the end result of that was about, you know, two years of R&D, where we ended up creating an analytical chemistry methodology that serves as our data set for our AI, that actually grabs all the understanding of analytes in a solution at once. And then it took some time after that to,

7:34you know, figure out the commercial application for the wine industry. But because we were able to acquire this unique data, we tied it to consumers and we focused on training our algorithms to predict the likability of products as opposed to what words they would use to describe their experience. And we can do that with 93% accuracy today. And our company tagline is we taught an AI how to taste and smell. And maybe for the more technical people out there, what we're essentially doing

8:08is we're creating digital twins of chemistry, right, whether or not that chemistry exists in the cloud. And we're creating digital twins of consumers. And we're running simulations to predict what is the best possible outcome or opportunity or white space for a wine. And so that way, depending on who you're talking to across the supply chain, we're really acting as a matchmaker between brands, retailers, distributors, consumers. So there have been many, many applications. And our challenge was figuring out

8:45where the most compelling use cases were. That's fascinating that I want to follow up later on your mention about digital twin, because, you know, Harvard Data Science Review actually just launched a special call for special issues on digital twin. But for now,

Data Collection and Quality

9:03I want to kind of dive into probably the most fundamental questions anytime we talk about data science, and for both of you, is the data collection and the data quality, right? Because that drives everything, I think, for both of you. So I'd like to hear from each of you, like, you know, what is your data collection process? How do you guarantee the quality? I know, you know, it's easy for me. I'm kind of a special expert on talking about data quality, you know, writing articles about the series, but practice is a different story, right? How do you deploy things? How do you make sure people follow what you tell them? So I'd like to hear from both of you

9:40how you ensure the data collection data, you know, quality in your practice? Maybe go with Kia first. Yeah, absolutely. So we understood data quality was very much a top priority. And part of what we're trying to do with our data is make it more accurate than a human walking around with a tape measure, trying to measure and count and remember what they saw and write it down on a piece of paper. By the way, the name Scout for our company is inspired by scouting, which is actually something that people

10:13do in the vineyards. Usually they're interns, they're given the task that nobody likes in the wine industry, which is, you know, go out for eight hours and count things instead of drinking wine. So in the beginning, we started calibrating ourselves on what is the human based error rate? And then how do we understand what the current as is, is, you know, first thing we did with our customers is we started asking them for their maps and their existing data. And I remember the early days was very, I was very

10:46nervous because 100% of our data was different than the customers. And what we found over the first year, 100%, 100% of our customers had, but this is the good news, Charlie, their data was wrong. And that was shocking to me. So what we found was, it's a target rich environment where we're comparing things that, and by the way, it wasn't the intern's fault. You know, I think in some cases, people tried to shortcut and assume that a vineyard is square, therefore it has the same number of plants. And what we found is

11:22that's not how farming works. People will not have the exact distance between the plants and sometimes two or five sneak into a row that are different than the neighborhood row. And so not everything's a metrics. I mean, a lot of people think of these as grids with the same number and they're not. So we feel right now, we've kind of seen this, we have 98 to 99% accuracy. The 1% or 1.5% is really distinguishing between a young vine and a weak vine, right? And that's a human problem as well.

11:58Most, but in terms of understanding whether something's in the ground, is it root stock or is it a mature plant? Is it producing or not? Those are areas that at this point, I would say we don't have a data quality problem. How we instituted this is within the data pipeline. And in fact, we have now a data czar, if you will, completely separate from the engineering team, whose job it is to literally look at data quality as a metric end to end. And that helps each of the

12:30pieces in the pipeline understand what their quality measures are and how do we assign a risk score to the data as it comes up so that the data itself in some ways is self-describing. So if I took a photo on a rainy day that was blurry, that should be rated as a lower quality, higher risk, you know, for error. So we're trying to ingrain this data quality concept, not just in a person or a team, but literally throughout the whole system. So we're actually using AI itself

13:01to also identify high risk potential blind spots so that we can stay away from it. The area in quality right now, data quality right now, that's very challenging and I don't, and this is physics and biology, is that obviously grapes could be behind a leaf. So you can't see visibly all the grapes. This becomes an interesting problem, especially for sparkling wine, where people don't like to thin any leaves and they'd like to leave it there so that

13:32it naturally comes out. And you'd be very happy to know that we're using a sampling algorithm to be able to now direct humans to go to certain spots and count the actual fruit and then use that to calibrate the AI models, you know, from that set. So in-ground sampling still is a very important component of our solution. And humans are always in the middle, either helping us calibrate our model better or, you know,

14:06humans who are doing the labeling or doing some of the other, you know, elements that right now cannot be done by AI. Yeah. So the, the core foundation of our company, Tastry, is built off the notion that our new invention, our means by which we acquire data that is unique, is what allows us to answer questions that were, traditional technologies were unable to answer. So from the very beginning, the company started

14:39because of this invention that allowed us to generate our own data in-house. So there's really two data sets that we have to generate in-house. I don't have, you know, historical data sets on past consumer purchase history or the Nielsen top 2000 wines sold in the US. That's all showing what happened in the past. What our data allows us to do is, in a much better way, anticipate the future. And so one data set is the chemistry. It's the analytical chemistry methodology. We have a lab based in the

15:18Central Coast. It's a third-party certified lab. It is compared against other labs, you know, in the country for accuracy and reproducibility and things and like that. But the key distinction is our methodology is much more efficient and, again, can grab all the compounds at once and the method is proprietary. So the lab is data set number one. Data set number two, which allows us to understand chemistry relative to consumers, is we have an experimental design where we first run a traditional focus group

15:58where we have consumers who would taste and rate the products like wine in the focus group. And we would ask them questions about how much they like those wines. Those wines we have the chemistry for. And once we have that data, consumers rating certain number of products, we introduce a third data set, which is we ask them questions about their preferences for foods and flavors that may or may not be correlated with wine. And I'll give you an example. You may have seen some gimmicky looking quizzes on

16:34on some wine websites in the early 2010s that ask you things like, do you like dark chocolate or black coffee or licorice or bell peppers? And the reason we introduced that data set is because we know that consumer preference is strongly tied to culture, cultural upbringing and what you grew up eating. And so once we have that data set, we can eliminate the need to have a focus group in the future just by understanding what that person's general palate is. It's not a one-to-one relationship,

17:11right? So I'm not saying, hey, if you hate bell peppers, then we're not, we're going to say you hate wines with pyrazines and it's more complicated than that. But it does give us analogs that give us a much more efficient, much more scalable way to generate that synthetic data and extrapolate what the other palates in the country are like, even if we don't have tasting data on them. So it's really the interplay of those three data sets that allow us to make predictions.

17:42I want to ask you guys how you guys are using AI outside of Scout and Tastry. And in Katerina's case,

Applying AI Outside of Scout and Tastry

17:50Tastry started with wine. I know, Katerina, you're working on other spaces now. I would love to hear from you on that. And then Kia, you have a project that's close to your heart, which is Neotempo. I would love for you to talk to the audience about that. So maybe we can start with Katerina. Maybe you can share how you're using this technology in other spaces and then we can go to Kia. Sure. So when we developed the IP of intellectual property, we had validated that our methodology applies to anything sensory based. So that could be coffee, it could be soft drinks, it could be laundry

18:24detergent, it could be fragrance. But you know, this was my first company. And I was just a scientist when I started this. I was not a CEO. I am now. I just wear a lab coat for photo shoots now. But back then, I was actually wearing a lab coat. And so we said, look, we need to focus on the vertical we were born in, which is wine, prove it out in this vertical. And then we'll expand and into other industries. And to be honest, growing in the wine vertical took so much longer than I expected.

18:57And we quickly realized we picked the hardest vertical from a technological perspective and from integrating AI into the workflow perspective. For example, the wine industry has what 160,000 SKUs coming out every year or that are registered with the TTB. That's not how many catch up labels there are. Second, it's an agricultural product. So we're picking up on nuances and terroir and all these things in the oak that could be influencing the seasonal effects on that. And then all the

19:31available chemistries and complexities of wine required a very large investment to acquire that data. And then thirdly, wine, as you know, is a very unique industry. It's fragmented. It's not the fastest adopter of technology. You know, there's a lot of artisanal practices there. And it's cyclical in the sense that they're during different parts of the season, you're dealing with different problems, right? You're dealing with harvest, then you're dealing with blending and all these things. So we

20:04were in the wine vertical only exclusively the wine vertical much longer than we expected. And we've done a really good job. We've worked with a lot of amazing winemakers, some of the most prolific winemakers in the industry. We've worked with, you know, some of the best brands, some of the most iconic smaller brands. But I just wanted to say we will continue to do that. But when I pop my head back up, and this has only been happening really the past year or so, and I started to talk to

20:35CPG companies outside of wine, I realized how much easier it is to implement this technology in other industries so much faster. And that the value propositions are a little different. So if I'm talking to, I don't know, a fragrance company, or, you know, a Coca Cola or someone like that, I'm telling them, hey, I can cut your focus group costs by 80% and help you bring a product to market

21:06in four months instead of 12, and flip the failure rate of that product from 80% to 20%. That's a pretty easy pitch. When I'm talking to the wine industry, it's a lot more nuanced than that. And depending on the kind of winery it is, the value propositions are completely different. What a small boutique cult winery needs is completely different than what a large production that is nationally distributed needs. Their problems are completely different. So we do continue to expand, but it has been

21:42very interesting to learn just how unique the wine industry really is. Yeah, this is amazing. Thank you, Katarina. Kia, tell us a little about Neotempo and how you're applying AI and, and as they would say, drinking your own champagne or eating your own dog food. Yeah, I think, I think starting our own brand and really focusing on farming first for 10 years, years. We grew grapes and sold the grapes into some of the best brands in Napa Valley,

22:14Shafer and Dariush and other brands really gave us a perspective around what does it take to make high quality grapes. And then in 2020, right before the pandemic was when we made a decision to move ahead with creating our own label and our own brand called Neotempo. We called it Neotempo because we wanted to be very proud to be a modern contemporary wine brand, not afraid of applying innovation from dirt to glass. So everything in our project is different. Almost every function is data

22:52driven and technology enabled. So we're the first to use Monarch tractors for entire vintage. We're on our third vintage of having electric tractors. We did the first autonomy demo of a tractor in Napa Valley in the back of our vineyard. We have a ecological sustainable model that is unprecedented in terms of capturing rainfall and recharging the well. We really kind of pushed sustainability and organic farming on one end, but then married it with technology and innovation in terms of irrigation sensors and

23:26having kind of automated stress-based basically water as opposed to regular water usage. We actually cut down on our usage of water by having more intelligence directed. And in fact, in the 2022 horrible heat waves, that technology saved our vineyard and then saved the vintage because we irrigated for seven hours from 4 a.m. to 11 a.m. before the peak heat came in the afternoon. The vines were hydrated and

23:56that's one of my personal favorite vintages in a very difficult vintage for other growers. We also are unique in that we share all our experiences. I have an open source mindset. Smartfarm.ai is a website we created. We put everything, the list of all the technologies, and quite frankly, Scout was born out of that experience. We've also taken the precision concept from not just in farming, but to winemaking. I was just mentioning to Catriona, I was one of her biggest fans of what she's doing in Tastry.

24:28We've applied Tastry. We find a lot of value in really comparing our wines to our wines. One of the things we've promised our consumers is that we make wines that are unique. And when you say that they have to be unique, they have to be unique even across a family. So what's very interesting, I'm sure Shaoli actually tasted two wines that came from one foot apart in a vineyard that chemically, they're very different. And Catriona's software can actually show that in Tastry. We can actually

24:59show the chemical differences, even though it's the same farming methods, same winemaker, same barrels. So we love showcasing small parcels and the power of exposing that terroir into the flavors. I couldn't agree more that the wine industry is very complex. I think with Scout, what we focused on is the number one hogs item in any winery, which is farming. And then how do we help them reduce this variability and

25:34complexity by at least being able to separate out the different parts of their vineyard and manage them differently rather than having everything kind of go into a soup that then, you know, you're trying to accommodate different palates and different flavors. So through the combination of AI and data, we can actually create three SKUs from one vineyard that tastes very differently, farm them organically, use precision to make them very pure and quite frankly, make them also profitable because instead of having one wine that

26:10is lesser than the three, we have three wines, all of which can basically perform from both the quality and price level. Other functions I use, AI, just, you know, my wife and I are the only, you know, full-time employees on, you know, Tempo. We use AI everywhere in our tasting room for translation of tasting notes to 12 different languages, all the way to music pairings, food pairing suggestions. One of the things we love is just using the concept of AI to recommend ideas. I have it almost as an

26:45advisor for the business. I have it give us ideas around placement. I was in Hawaii on a trip and literally went through a long prompt to honestly look at distributors and restaurants and suggestions on restaurant components. It's incredible what you can do. As long as it knows your brand and knows your preferences, it can be really used to do a lot of good. Since this is a data science podcast, I'm going to go back to a little bit of nitty gritty in terms of the next step. You guys talk about

27:20collecting data and the data quality. I want to ask both of you, after you have the data, there's lots of, you know, database management, there's a data analysis. Can you speak a little bit about, you know, what do you do? Are you using traditional, I don't know, regression methods or you're doing deep learning or something fancier? And it's a combination of both. And how do you make sure your database is working properly? You know, things get changed, different versions. You know, I want to be slightly nitty gritty, but for this audience, like in practice, how do you deal with

27:53all these seemingly mundane, but they could also change your conclusions? Yeah, absolutely. So we collect all of our data in a time series so that we can, first of all, we've kind of thought about this as something that over time provides a new data set. The changes, because we're looking at the same vineyards season after season at different times of the season, the data set that says what has changed is also very, very important and in effect becomes a new data set. We use a number of,

28:27at this point, pretty standard techniques around regression, anomaly detection. Some of the forecasting and prediction is very interesting. So for example, in yield forecasting, we ask the clients to upload their historical yields. If they have those, those become some boundaries and some hints, if you will, for the model to know whether it's in line or if there's massive discrepancies. And if there's discrepancies, there better be a reason. So we're almost going to use that as another guardrail for data quality to give us what normalized limits are. I think one of the big challenges that we have

29:03is a lot of the data that exists today and historical data are manually captured or in Excel spreadsheets. And we've now began to ask customers to just give us their historical Excel spreadsheets and we're using RAG and other techniques to kind of bring that data set in and store it in a more structured fashion. And then the last part that's also important on the AI side for us is that we introduced a new product called VIA earlier this month. VIA stands for the Vineyard Intelligent

29:36Assistant. It's effectively a LLM extension as a chat bot that plugs right into ChatGPT or Gemini or Claude. And it has a database of about a thousand articles. So if you think about traditional scout as we're taking private data and indexing it for the vineyards, VIA indexes the public trusted reference data that effectively would be the equivalent of a massive library on viticulture best practices and

30:06field notes. What I'm excited about, and this is on our roadmap in the future, is bringing both of those data sets together. Because then imagine a world that you're capturing these photos and the metrics and the vital signs. And at the same time, you're layering on top of that all the best practices. That allows us to become very proactive in telling people things that we're seeing without having to wait for the user to ask us. So that's kind of a little bit of the detail. I can, you know,

30:37we're on Google Cloud. I can go into, you know, kind of the different layers of the stack. But we try to, to the degree possible, use horizontal generic technology that's proven, and then invest our core IP on viticulture or agriculture-specific models and domains. Katerina, same question for you. Yes. And I will be speaking as a chemist and on behalf of my PhD data scientists and mathematicians. So just bear with me. So there's two things. One is there was an AI or machine learning model, I should

31:14say, that had to be developed in-house to work with our data. It was an invention in its own right. There was not a term for the kind of model it was, I think, when we did it. I think the closest thing that exists out there today is maybe like a liquid neural net maybe is one way to kind of think about it intuitively. But essentially, we had to find a unique way to associate disparate data sets in a

31:44multi-dimensional space, just to give you an idea. And so that's a really cool piece of our secret sauce. Now, there are many other data science techniques and AIs we use beyond that to extrapolate information that we gather. You know, one example is when we're trying to create a heat map of real and synthetic data on consumer preferences in the US, we have to do like a Bayesian ridge regression, right? Or some

32:16Bayesian ridge statistics to extrapolate what is the distribution of pallets on a store, local, regional level. So there's all those other techniques that we have. And that's kind of like a living, breathing thing that is updated continuously, right? Because we have new products coming out every year, new pallets coming in every month. And so you can see like the drift of preference versus products over time, but you could always go back and see what it was or try to go

32:51forward and see what it will be. So Katarina and Kia, this has been a great conversation so far.

Opportunities and Risks of AI

32:59There's a lot of anxiety around AI. And rightly so that we were hearing layoffs, we're hearing a lot of changes. A lot of the layoffs might not be related to technology or AI, but still some might. And there's people who think that AI is going to take away jobs. There's people that say that no, there might be jobs taken away, but jobs will be created as well. Others think that it's not really the jobs that will be taken away. It's more like certain tasks within roles will be start being outsourced to AI. But at the same time, there's a lot of opportunity and hope and a lot of people

33:32talking about the wonders and how AI can help solve things. What are the biggest opportunities and what are the biggest risks that you guys see that AI has? Maybe you guys can comment on that, Kia and then Katarina. Yeah. So the first opportunity I see is unfortunately, the wine business is one of the most inefficient industries. It's a multi-billion dollar industry. It's under pressure on margins. So I think without AI, there will be a lot of layoffs because it's not a well-running, efficient business. And I think there's

34:07hope with AI that some of the big problems that are nobody's fault, by the way, it's not like people get out of their bed saying, hey, let's go work at an inefficient business. Let me give you an example. It's the only multi-billion dollar business that I know of or industry that I know of that doesn't have a forecasting model until the fruit is literally picked very late in the season. That means that every winery in the world has to make guesstimates on how many bottles and how many

34:38packages and how many corks and how many labels they need to order without having the actual data. And so if we're targeting yield forecasting at a block level to be accurate within 9%, the current rate when we survey our customers is somewhere between 20 to 30% inaccuracy. That's unimaginable. Can you imagine a car industry where you would buy a bunch of tires and not know how many cars you would manufacture, and then you'd pay for a bunch of warehouses

35:12where these tires would be stored for the next season. So I'm very optimistic around AI. I think what it will do is cut waste, cut tasks that nobody signed up for. As I said before, if you're an intern and you join a winery, you didn't do that to go do manual vine counts. I think it will bring young people in. We've seen that with technology shifts. I want to remind everybody, we went from horses to tractors. Tractors did the job much faster. We went from landlines to cell phones. And again,

35:44nobody's complaining other than nostalgia about landlines. And one thing we know is we can't stop technology because in this type of world, the people will learn how to make it work. Humans will learn how to work with technology. It absolutely has negative consequences. I'm not in denial on that. But I think if we're pragmatic and we don't fear monger and we don't make it sound like, you know, everything's fine without technology, we will get through it. And maybe on the other side,

36:17what we will have is a more efficient, better performing industry with a workforce that is trained on data and knowledge. And we let the AI do the things that, quite frankly, humans are not great at, which is computational models that are four-dimensional. I mean, the reason at the beginning of this podcast, we talked about wine is wine is one of the most complex subjects that exists. There's so many variabilities. You know, if you look at the cardinality of all the things that can contribute to what ultimately is in your taste, it goes from soil to

36:52biological matter, rootstocks, clones, you know, weather, temperature, picking time, you know, bricks, you know, phenology, chemistry. So I think AI is phenomenal for industries that were underserved by old tech. None of my clients have a large research budget and internal developers and data scientists that are sitting around building apps. So they can kind of almost skip that entire

37:22generation of tech that required expensive developers and expensive data scientists, and maybe adopt technologies that were built on AI that are smarter, that they give the end experience that they need without having to spend millions of dollars building this in-house. I 100% agree with everything Kia just said. So I would speak from my personal experience. You know, trying to sell a technology company into the wine industry took a lot of annealing the market,

37:54if you will. And when we first launched, the number one kind of joking kind of feedback I would get is winemakers would say, Oh, so you're going to replace my job. And, and it's ironic because it actually turned out to be the opposite. It actually saved a lot of jobs because the reason they took a chance to use us initially was to solve emergencies and disasters that were happening in the production process. Like, like when this fires, we had the crazy smoke taint season and we had to figure out

38:29how to ameliorate tens of millions of dollars of crops or, or wine, or you were going to lose it. We were able to work with the teams to solve those emergencies that would not have manually been possible to do fast enough, you know, the way it's traditionally done. So that's actually how we got our foot in the door is solving problems. The winemakers just didn't have the time or resources to do. And that established a lot of trust. And I think once they were able to get their hands on

39:01the platform, you could see that it's really the human that is guiding and in charge of the AI. It is, it's your slave, not the other way around. You still need context, like you still need inputs. And it's really like a paintbrush for the artist to be able to evaluate options more efficiently. And I just don't see for AI like this, I don't see the future doesn't have a human in the

39:32loop. I don't see that getting adopted anytime soon. It's really not AI that's replacing humans. Honestly, I see it in data science and software development and design. It's really humans using AI that are replacing humans. Yeah, that's great. Thank you, Karina. Chely? Clearly, this conversation can go on for a long time. And I, you know, Kia, you're absolutely correct. I, you know, that's a really main reason that we choose wine as a feature topic

40:03for Harvard Data Science Review, other than, of course, I love wine. But there are so many factors. We haven't even talked about all the other factors, you know, how the wine was served, how long it was decanted, who are you drinking with, you know, all these factors all affects, you know, your enjoyment, right? So, but we do need to wrap up this time, because, you know, it's getting to time. I like to have a glass. So I like to complete this wonderful podcast, but we always complete with a magical wand question. So the magical wand question for both of you is that if you can wave your magical wand, what is the new AI technology you'd like to

40:40have for what you do? For Tastry? For Tastry, yeah. For anything you do, like what would be something you don't have, but you wish you have? I would love to expand into other verticals and have the resources to do that. So I would love to move into the other categories that are interesting to consumers, right? There's, there's so many other, you know, I know cannabis or CBD beverages are of interest, non-alcoholic beverages, the ready to drink category is exploding.

41:13Um, and I'd like to see how that correlates to those particular consumers' wine preferences and how it, um, how you can guide them down a chain of different products. So more data, more data in different verticals is my, it's going to have to be my answer. That's, that's absolutely important, but what would be the new AI technology you would need for doing that or the current existing ones is, is can serve your purpose? Well, it would be, I'm still speaking a little bit as a scientist, but the aspiration for what

41:49the new technology would be, it was to start correlating a one single consumer's preferences across many other verticals. And what I mean by that is, if I know that you like these oaky, buttery chardonnays, for example, does that mean that I can predict that you're going to like the butter pecan ice cream on the shelf, which is a completely different product? Thank you. Kia? Yeah. So I would say, um, if I could have a magic wand and had, I had the data set, uh, you know,

42:25that, that, that we're building, I would love to be able to get to, uh, a forecasting model that could combine forecasting and what we see on the ground with weather. And this is something we've been looking at, um, very closely. So it might not be obvious, but grapes shrink, uh, during heat, unlike, uh, apples and oranges, and that has a massive impact, uh, in terms of weight. So you could have a heat wave that you lose 20% of your water weight and being able to marry these two very

43:00interesting time series data that potentially happens in the future and being able to build some calibration between them so that there's a correlation. We know, for example, that at 20 bricks, uh, grapes shrink, that's the maximum peak. And then they start going on certain varietals. Again, a lot of this is research, but in the world of AI, you take that raw basic principles and you start building models and, uh, ways that you can calibrate it because we don't know what the future holds. We know what the past holds. You know, one of my favorite things is taking hindsight and

43:36turning into foresight. And so we love, uh, looking back at the past as a indication of the future. In fact, one of the exciting new projects that my, uh, co-founder CTO Mason's working on is this notion that like literally on every day of a growing season, he's narrowing down. What are the scenarios? What are the previous vintages that were at? And then when you get to the latest part of the season, either it's a new season on its own, or you've reduced the variation down to the

44:08scenario. So it's kind of an interesting way of even thinking about AI, um, not just in having an answer today, but becoming a much better forecaster of what could happen. And that would be something that if we had that today with the data underneath it, it would make a huge difference.

44:29Thank you so much. That was just wonderful. And, uh, I want to thank all three of you for your great contribution and thank you so much. Thank you. Thank you so much. Thank you.

44:44Thank you for listening to this month's episode of the Harvard Data Science Review podcast. If you are interested in the Vine to Mind, next year's conference will be at UC Davis from May 18th to 21st. Please look out the website vintomind.org for future announcements. To stay updated with all things HDSR, you can visit our website at hdsr.mitpress.mit.edu, or follow us on Twitter or Instagram at the HDSR.

45:17A special thanks to our executive producer Rebecca McLeod and producer Tina Tobey-Mac. If you like this episode, please leave us a review on Spotify, Apple, or wherever you get your podcasts. This has been the Harvard Data Science Review, everything data science and data science for everyone.