Blue-skying machine learning in retail and fast-food (Ep.78)

Episode 63 · 50:44 · SRT · archive.org · all Tanzu Talk

All right. Well, we were I was hanging out in Twitter the other day. Obviously, you were. And I noticed that you you tweeted out a as you called it, a architecture diagram, which one of my most cherished favorite terms. That's a good one. It's it shows a word like architecture. It's it shows multiple layers of ironicness and sort of like tech cultural deft to just kind of throw that about. So that's that's that's a delightful one. But I got, you know, as as as you and I were talking, mediated by the computers and various channels that we mess around in, it seemed like it'd be good to have you on to to talk about a little bit. So why don't you introduce yourself first? Thank you, Coté. I should say longtime listener, first time caller to be here. This is this is definitely a treat. So my name is Scott Truitt. I've been at Pivotal since I guess I joined the week we became Pivotal, which is still surreal to think about. And I've played a number of roles in the Cloud Foundry team. Most most of them have been product related. And that really is my background. I've always done product. I've always worked in software startups. I've made hundreds of these architecture diagrams. And I agree. I think it's one of the best words we have to describe what we do. It's at once dismissive and celebratory. Exactly. And I think that's that's kind of an awesome thing for us. It's kind of like this is going to be one of my patented terrible analogies. But if we if we imagine that blowing coal smoke into the air wasn't actually bad and it was OK, it's the equivalent of like those people who are those diesel people who fit a big pipe on their truck. And they're like celebrating burning diesel fuel. And I think I think and it also powers their gigantic absurd truck. But I think that's kind of what a architecture diagram it's like, you know. Yeah, you're at once broadcasting that you're not technical, but you're doing something that's deeply like in the weeds on what the product really is you're building. That's right. It's a regular. I see what you did there situation. Exactly. So so this diagram is not I'm looking at it again.

I don't think it's labeling what the application actually is. But and I'll put a link to this in the show notes just so people can see it. But, you know, what's what's interesting about it. So let me give a bunch of context. I mean, if you're a longtime listener, you realize that I spend about 80 percent of the time. That's the best part. That's right. I just did that with with. Let's see, probably the order I published this in with someone I hadn't met before, one of our colleagues in the previous episode. And I was thinking, like, man, I hope she doesn't think I'm a jerk. But anyways, so you've got you've got what's good about this is so I've been over rotating on figuring out what enterprise architects do in a cloud native world, what they should be doing to help out and what what they're a positive, good role for them is how they can be. I'm going to stop using this after this, how they can be the big old pipe on the diesel truck, as it was. But man, I saw one of those the other day and I was like, what are you thinking? To be fair, you live in Texas. I know, I know. Yeah, I know. So, I mean, it's odd that it's only random. Well, it doesn't it doesn't it doesn't show up in Austin very much. We went camping like out in the country. And so, you know, there you go. You go back to the I don't know. I hesitate to call it real Texas, but the the the Texas most people think of. Yeah, let me put it that way. So anyways, I've been trying to figure it out. And I think, you know, one of the key things that enterprise architects, one of the outputs they have, so to speak, is an enterprise architecture and dropping my silly analogies, which actually, if done even averagely well, is actually incredibly helpful. Right. Because it agree communicates how everything fits together and explains the different components. I mean, it's it's it's a map, not to be all wordly. I don't I haven't read his stuff enough, but it's a good map of what's going on. And so you tweeted this one that has a good swim lane that talks about how data is moving through, how various data processing things are used and platform and also how like TensorFlow is used and all that kind of stuff.

And so I thought it'd be good to just have you come on and talk about that. And also, it's a good way to highlight the actual how actual technologies that we go on and on about at Pivotal kind of fit into scenarios like this and integrate together. So with that long introduction, why don't you tell us like what this is a diagram of and kind of what it's explaining and just kind of we'll start going through it. Sure. Maybe maybe I'll take a step back and talk about the genesis of this project. And that was really a desire to go talk to retailers. And, you know, we thought that was an interesting vertical. Certainly, there are competitive pressures from a certain company in Seattle that make retail ripe for disruption. And when we started it at the end of last year, no one had any inkling that Amazon would come in and swoop up Whole Foods. Yeah. So, you know, it was it was somewhat prescient. But either way, we got into this. We realized that maybe the most interesting thing we could do would be to act like a retailer and to be forced to make decisions around, well, what can we do? Where do we focus? How do we get closer to our customers? It was about this time that I saw a great quote from Jeff Bezos that said something like customer service is our core competitive advantage. And we think we can win that battle each and every time. Poor paraphrase, but I don't have it in front of me. And to me, that isn't something to run from. That's basically that's a call to arms. Let's go. Let's do this. Right. And so we kind of scraped together this notion of, OK, if we're going to say customer service or customer experience is a competitive advantage, what would we do? Right. And we thought about, you know, our kind of history at Pivotal, the companies we've worked with. We, you know, I'm familiar with Gap living in San Francisco and understand their kind of, you know, their situation, for lack of a better word. And we imagined what it would be like to be a clothing retailer in this space.

Of course, we don't have access to all their info. We certainly don't have their data. So we went out and made one up. You know, it's loosely based on them, but it's pretty generic. We scraped another retail site out of the UK to get an inventory. And then we created, you know, a Twitter account. We also created a user named Maya who had a preexisting relationship with this retailer and was just benignly tweeting about a party she was going to this weekend with her best friend and tweeted this picture of a dress and said, I wish I had time to go buy this. Right. So the machine we built behind the scenes pulled that tweet in, looked at the content of the words, put it through this natural language processing API service that comes from Google Cloud Platform. So we understand the sentiment. We understand the words in her tweet. We also use the Google Cloud Vision API. And so there now we understand the image. And at this point, we have a really interesting piece of the puzzle. So we can then take that, compare that against our inventory. We can run a machine learning algorithm against it built on TensorFlow. And we can determine if we have an item in our catalog that matches her needs. Even better, we could match that up with a store that's close to her and send her a code to save 30% off. And we can do that all in about 10 seconds, which is really quite amazing, I think. So that was the gist of it, right? We touched a bunch of different things along the way. But the fact that we were able to make this magical thing happen, completely benign tweet, turns into an offer that's personalized specifically for her with a time, date, and location. That's pretty compelling stuff. Yeah. No, that's a really, I mean, there's all sorts of things to sort of bite into there. Not to mention the thing itself. Right. But, yeah, so first of all, it's interesting.

I think the phrase, when you were first saying it, I was a little confused. But the phrase customer service, I guess not to be too insidery, but it's one of these terms like CRM, customer relationship management. Yes. It sounds, or, well, I won't use that third one. It sounds awesome. And then you see what it actually is, and you're like, oh, it's more or less the opposite of what it's supposed to be, like in practice. And so, you know, when I think of customer service, I think of basically, I mean, I first think of, you know, being nice to the customer, right? But then I kind of move on to, yeah, yeah. But then I pretty much, I mean, there's other stuff, but more or less I move on to, like, where did I put that receipt for this thing I need to return? Right? Like, or, you know, like, because that's, I feel like the core, I mean, that's literally usually called the customer service desk. And so I was, that's why I was a little confused, is I was thinking, like, I was just telling my wife the other day that the one thing that Amazon is annoying about is returning stuff. And so, like, so thus my head was clunking a little bit. But then as you talked about it, it's sort of returning to a more genuine, literal idea that is, it's more like being Jeeves in Mooster and Jeeves. It's anticipating what the customer wants before they want it. And more importantly, trying to be helpful. And of course, the reward, just like I'm sure Jeeves was compensated well, the reward is that you get paid, right? And yeah, I think that notion is good because, you know, I won't lie. There's often times where about 40 to 80% when I complained about some brand on Twitter is I'm secretly hoping they'll give me something. And it is, so then the other thing before we get to the actual thing is I think, I don't know, tell me if I'm over overthinking it. But I think this is potentially a good, practical, pragmatic example of, I don't, I think we need to have a new phrase for this.

But whatever people actually mean when they talk about AI and machine learning nowadays, right? Like it's not going to like drive a car for me or have a conversation about the true meaning of Nietzsche with me or anything like a science fiction AI. Yeah, exactly. But, but it is kind of like, and this is like, I don't know what to call this, but what you're describing is sort of like, what are, what are advances in computation, right? Computers are faster and cheaper is one thing. And then I assume we also have some new ways of telling computers to find patterns and doing things automatically, kind of like machine learning. I'm grossly simplifying, but like. But I think you're right. What are some things we could do just to like do new types of business, right? Yeah. And so it seems like a good example of that. You know, I even think back to the, to the first stories I heard Paul talk about in the early days of Pivotal, and they were really, truly data-driven personalization. It was understanding the context at that moment and making the best possible decision for both parties, for us as a company, for the customer, as our constituent. And I think that's a really interesting territory to mine. I just think there's so many things you could do. And of course, we barely scratched the surface, right? We made up this woman. We made up this retailer. We didn't have access to a click stream. We didn't have access to past purchases. We didn't know if she liked stripes or solids or patterns or prints or whatever, right? We were basically going off 140 characters and an image and doing something magical with it. And I think that is, is really what I realized the most compelling thing. We ran an experiment. We saw interesting results almost immediately. I mean, this whole thing took on the order of six weeks for a handful of people to do. That's not a lot at all. No. And when you think about, we didn't have to deploy infrastructure, right? We didn't have to go create a visual image processing service that we could use. We simply did a CF create service and a CF bind service, and we used it, right?

It's all in the Google Cloud Platform service broker that works on PCF. Same with the natural language processing API. The TensorFlow stuff, we had some data scientists help us within Pivotal, but this was in their wheelhouse as well. And then at that point, the real shining piece for Pivotal was Spring Cloud Dataflow. And we had this ready-made platform, or I should say pipeline, to bring raw data in and help transform that into insight and action. There was very little, it's not fair to say custom development, because certainly the whole thing was custom, but we weren't building at an infrastructure layer. We were building interesting things on top of what was already there. Yeah, so let's start with that pipe. I mean, I feel like somewhere in this podcast, we've talked about Spring Cloud Dataflow, but like how, so what is that? How would you describe what it is? Let's see, here again, I'm product, so how it works, I just know what it can do. That's fine. We can be all like benefits and outcomes, and just kind of give a little rub on the fat belly of the value prop, and we'll be fine. Sure, sure. What do you do with it? When you look at that architecture diagram, there's really four main pieces to it. There's the initial sync, where we bring the data in. And then as we push it down the line, we do some processing. In the second column, that's where we do the interrogation around the language and the image. When we get to the fourth column, that's where we're doing the just cold, hard data science. Do we have a match? What's the best match? What are the four best matches in our inventory? And then in the fifth column, we're really taking and batching that output and sending it off for action. So it's this notion of, you know, in my mind, it's glue code. It's incredibly elegant and very simple to use.

And the fact that we could do this transformation, but that is our backbone, proved to be really powerful stuff. Yeah. Yeah. I mean, it seems to me each time one day I need to actually sit down for like maybe three hours or longer and like just kind of delve into spring cloud data flow more. But it seems like, I don't know, one of my favorite boring types of enterprise applications is like the sort of workflow with data transformation stuff. Your classic ETL. Yeah, yeah, yeah. And to your point, right? Like the goal of a lot of these things is we want to we have a decision process that has some input to it. And we want to do something more than glue code, right? Like we want a formalized way of doing it. And then, of course, like in Pivotal Think Land, once you formalize and standardize it, it means you can manage it extremely well and automate it. And then to your point, you can do it very quickly and reliably, right? Yes. So that's always nice. But it's essentially a workflow. And then I think the, I don't know, maybe the thing that I don't know if it makes it boring for people who care more about, you know, hot dogs and stuff is like the end goal is someone is sitting there and makes a decision, right? Right. It's all just sort of augmenting a person's decision making capability at some point. Yeah. In this case, we coded that decision into it. But certainly you could go down that path as well. Yeah. Yeah. And that is, it's also kind of a hopeful thing for like the robots are going to steal our jobs is like, you know. Maybe, maybe for all the people who are saying that they want like a nice pair of shorts, you have roomfuls of people at home who are just like deciding like, well, we should give them some shorts. Right. We're fitting into that train. Awesome. So, so, so like how many people did you have working on this, this thing? There were two of us at the core. As I mentioned earlier, Mike Goddard was, was the main developer, a good friend of mine, and just amazing ability to pull all this stuff together and make sense of it and build it.

And then we had some people float in and out on the data science side. We had some designers early on, kind of help us out. But, you know, at the core, it was a handful of people. And, you know, I think that's, that's even more interesting. We can also talk about the fact that we have a mix of Python apps and Spring Boot apps. We've got this, you know, Polygon platform where we could bring other people in to help us with very specific things. We had someone do an Angular front end so that we could demo kind of the data transformation process as well. All this stuff was easy. We didn't have to worry about it because that's PCF. And, you know, I think maybe the most interesting piece is that this thing took six weeks. How much time do you think our next effort is going to take? It's probably going to be three weeks or two weeks, right? The number of experiments we can run with this technology is going to increase as our familiarity and comfort with it increases. And I think that's really the story that resonates when I tell people. It's like, hey, go learn this stuff. But now go use it, right? Go free up your smartest people and tell them to go make mistakes. Go fail. Go do things that don't work because you're going to learn so much. Yeah. Next, y'all should do a rev where you scan all the IT people and you say, like, you might be interested in this white paper. That would be great. You could be your... I can't help you with that one, Coté. You could be your inside sales augmentation. Sure, sure. But, yeah, so, you know, it also strikes me that, like, you could use this as a good illustration of, like, how you would break microservices up, right? Whether you want to think of them components or services. Oh, sure. But to your point, like, I mean, I don't know, the various steps and components could be driven by different services and even teams that you might have. Right. Which is, it's always hard to find intelligible examples of microservices. Yeah. I mean, we did break everything up into microservices. Each one of those buckets is an app or a service.

So, that part is definitely interesting as well. There's maybe one other anecdote that I love to tell here, and that is the first time I showed it to a friend of mine at Google. And he said, okay, great, great. How are we going to scale this thing to a million users? Right. We don't even have a user, and we really didn't build it for that purpose. But here again, as I thought about that, I realized, well, it's the same infrastructure, right? I'm just scaling my apps, and I'm scaling my, you know, my knowledge of the system as it grows. But I'm not switching this thing out. I don't need to all of a sudden go to production-grade infrastructure. We built this on GCP, right? And PCF is production-grade. We know all this stuff can withstand that. So, what's a million users? Yeah. It's probably trivial. Yeah. It's trivial, right? You don't have to worry about scaling the infrastructure, just making sure your own code is scalable. Right. Yeah. Maybe our code changes. Maybe components get switched out. But those are solvable problems. Yeah. Those are a lot easier. So, then, as you mentioned, like, you used a lot of Google services in there. And, you know, for those who don't know, in the retail space, it's considered polite to use Google instead of Amazon. So, people are not so into using AWS. I hear Walmart has opinions. Yes, yes, indeed. So, you know, my friends in Redmond have compelling offerings here as well. It just so happened that we did this project with Google. That's right. And so, like, what components did you use from them? And what made you choose them other than the corny reason I was just joking about? No, I think it's a good question. As we looked at it initially, you know, about the time that we conceived of this project, the Spanner was announced. And we're like, oh, let's use that, right? We went from thinking, let's just show the art of the possible to, no, let's do state of the art, right? And we realized that Spanner was overkill for our simple use case here.

This is not designed as a product. This isn't something that we say, oh, here, just go run it, right? It's an illustration of what's possible. So we took some, you know, we didn't take the biggest leap that we could. We made some reasonable assumptions and we got in the game. So we used Google Cloud Storage instead of Spanner. But we still use TensorFlow. We still use the NLP API and the Vision API. We used RabbitMQ. We could have used Google PubSub, right? Some of these decisions we made were comfort and familiarity for us. But again, as we mentioned, those are services that you could swap out if or when that becomes a need. Sorry, I was trying to find the mute button. Now I'll have to edit something. Let me see where I was. 21 minutes. It's terrible. But it gives me, you know, makes me feel like I'm actually earning my job. Pause around 21, 15. So, you know, we're both product people, but I'm asking the question. So we'll see if you can come up with the answer. But so what does TensorFlow do? I had a chance to ask, you know, our buddy Schaefer about that the other day, but I skipped on it. But like, how does that fit into the overall flow of things? So I'd love to hear him describe it. I've heard my friend Jay Marshall give a high level overview. I think where TensorFlow is today is very early stages. But it has some pre-trained models. In theory, you could plug in and use one of those out of the box and get some reasonable output from it. I think what's interesting is when you can train up your own model, maybe building off one of those initial ones or coding one up by yourself. What it can do is probably way more than I can understand. But the way that we used it was simply, given these inputs, what are the matching outputs, right?

So we took our 4,000 items of women's clothing, and we ran it through the Vision API so that we had all the keywords on those things, right? We understood it. We could do a simple comparison to say, here's our batch of keywords, which items are our closest match? So in this case, it was a real-time kind of, you know, data science problem-solving machine for us. And it just so happened that we were able to bind to it as a service and use it in the context of this thing. We didn't have to manage our own TensorFlow. We didn't have to go and deploy it and configure it and all that stuff. It was simple, you know, again, create service, bind service, and then use it. And so essentially, so the overall use case, just to reiterate it for this, was someone writes a tweet that, like, I'd like to buy this, I don't know, tartan-colored skirt or something. I guess tartan's a pattern, not a color, but never mind. And the idea is you're going to proactively be looking for this, and you would suggest, like, oh, here's something in the inventory. And, you know, also, if you go down the street, you can get it for whatever, deals. And so to running that through, you know, no pun intended, the TensorFlow thing, it's sort of like you would take the inventory that you had, and let's see if I can not embarrass myself. You sort of use the magic of TensorFlow and whatever that is to kind of discover the nature of all the stuff that you have, whether it's through image recognition or looking over the metadata, and you're like, here is what we have, right, and the keywords. And so this gives us sort of like a model that you can stick things in and get an answer out the other end. And so if the thing we're sticking into this model is I would like a tartan skirt, it can figure out if we have something to recommend there, right? If there's something similar to that or something that is clustered around it or whatever.

Exactly. And, you know, that example, I'm sure on purpose is relatively simple, but it could also be something like, you know, I flew to San Francisco, and once again, I forgot that it's cold here, and now I'm chilly. Right. And, you know, this would be much more advancing, but it could be like, hey, dummy, you should go down the street and buy a jacket. Here's a link. Right, right. But, yeah, I think if your description doesn't show up on Google's marketing site by early next week, I'll be shocked because I think you nailed it. Right, right. But I think it is maybe this kind of all-purpose engine for comparing large amounts of data and deriving the best possible answer. At least that's how we used it. I'm sure there's a multitude of uses that I haven't imagined or can't even imagine. Yeah, yeah. But I think that's, you know, one of the things that comes out of that is, well, let's go try something else, right? We understand how TensorFlow works. Let's try a different model, or let's throw more variables into the mix, or who knows what. Yeah, you know, I probably will never understand all the, like, math behind it, but, I mean, there's two things. One, like, I think maybe last year I spent a long time, I don't know if you've ever read the DevOps report or reports. No. But I spent a long time trying to understand, like, the methodology they have. I mean, it's some sort of, I wish I remembered all the stat nerd terms, but, you know, there's, like, some sort of three Eastern European hyphenated name model for, like, cluster analysis. Right, right. And doing this thing, and then a bunch of, like, exponents. And, you know, Nicole, who's one of the main, or I guess the main academic person behind it, she very kindly has offered to explain it to me many times, but I've never followed up on it. But all I could figure out is that basically whatever all this funny magic is, is that it, the way I kind of filed it away in my head so I was comfortable with it, is it, it discovers, like, patterns in some data set that, like, didn't, weren't there.

Or, I mean, they were there, but, like, you, you weren't explicitly looking for them, right? Like, it says, like, these things are all similar, and part of what I figured out are the attributes to figure out how to figure out if they're similar. And then, and then ongoing, that kind of snowballs, and I get better and better about it. And then the second thing, and I, maybe they use this, but, you know, I switched over to using Google Photos, and it's kind of, like, with the shallow knowledge I have of computer stuff, it's fascinating to look at, like, the clustering that it does of photos, right? Like, oh, yeah. And it's very, and I guess App Photo Moments does this, too, but it'll be, like, here is three hours from two weeks ago that you spent in this part of town. Right. Right, and, and, like, you know, sometimes, like, the best ones will be, like, days where I have, I've left on a trip, and so it'll say, like, your day in Kansas City. And then it'll start off where I'm spending time with my kids for lunch, and then all of a sudden I'm in Kansas City. And, like, obviously that's a little screwy, but it's pretty good at, like, figuring out what are photo albums that we could put together for you that, like, no one would really think of doing. Right. And that's, that's what seems exciting about that. And to me, what makes, like, this kind of stuff more practical is, like, and again, you get it in front of a human. All we're trying to do is, like, suggest new stuff for you to do, right? Right. Like, you should tell this person to buy this skirt, which I think is interesting. And I got to assume that Google is training that model. Right. With everything that you do with it, or that it does for you. And I think that's the interesting piece as well. This thing gets better with more data, with more time. You know, today we may be able to make, you know, a broad match. But you can imagine if we knew the history, your size preferences, how often you shop, do you buy on sale or full price, et cetera, et cetera, and really start to craft a very personalized offer for you. I had the silly statement early on where I said, look, I hate ads, but I love deals.

And as soon as it feels like a deal for me, that's compelling. If it's an email blast about something that I don't care about, I'm less likely to move on that. You think about how much money is spent with just broadcast advertising versus what's possible with very pinpointed, data-driven, personalized deals. I think it's a huge opportunity for retailers to figure out where this thing goes. Yeah, no, I think that's precisely it with ads, right? Like, it's kind of like my quiff about legacy software, right? Like, you only call something legacy if you don't like it. Otherwise, it's just your software. And, you know, it's almost like you almost call something an ad if you don't like it. Otherwise, it's a good deal, right? Like, it's something that you enjoy. But no, that's an excellent way of thinking about it. And I think it's – I mean, it does show you, like, again, the practical benefit you can get from doing these things. I mean, there's always the now infamous folklore of, like, you know, telling some dad that his teenager is pregnant. So you've got to avoid that stuff. But – Right. I mean, you know, there's that kind of stuff. But it is nice to craft those things together more. So other than me sort of, like, you know, meandering through this kind of stuff, like, what's the actual real experience of, like, working with a data scientist? Like, how does that – how does that fit into, like, the balanced team, product team sort of thing? How do they kind of – do they just, like, write down some Greek figures on a Post-it note and put it up on the board? Or, like, what's that interaction like? I mean, I've worked with developers for 20 years now. I hate to date myself, but that's probably true. And I still believe what they do is indistinguishable from magic. Right. I understand what I want done. And one of the wonderful things about Pivotal is that I don't necessarily have to care how it's done unless how is going to impact my when or my why.

But the fact that we could, you know, almost monosyllabic grunt out some instructions to data scientists and then come back with this amazing model that returns the results we're looking for is pretty magical in my book. Yeah. So I'm assured that it's actual math and that there's no hocus pocus behind the scenes. But it does seem pretty darn magical to me. So the interaction would just kind of be, like, as you're saying, tell them, like, we would like to do this. See you later. And they immediately said, oh, yeah, well, that's this problem. That's well-known in data science. Here's how you do it. Right. Yeah, yeah. My result was two thumbs up. I'll stand over here and let you do that. Yeah, that's interesting because it's, I mean, obviously you would expect this from being, like, a pivotal thing. But it's a similar sort of part of the job of a, let's call it pivotal way engineer, technical person, is to not only know how to solve the problem, but know how to talk with someone like yourself or worse, myself, and be like, I think this is what you're asking. You know, like, I understand what you're saying, and I can translate that into my wizardry, right? Whereas I think, you know, the stereotypical, like, you know, beardy weirdo in sandals will be like, you know, they'll be like a command prompt. They'll be like, I don't understand what you've just asked me to do. Try harder. Type it in here. Right? Like, go read a man page. And I think, you know, even the word data science is cool, but it is a bit intimidating, right? Because it's sort of like, what, can I talk to these people? But, yeah, yeah. I mean, and that does, this kind of thing does highlight. So how involved were they in that six weeks? Like, did you, like, were they working with you constantly or did you just work with them a little bit? Because it also strikes me that something like a data science role is one of these not fully dedicated to the team things, unless, you know, you're in petroleum engineering or something.

But you might just use them on a consultative basis on the team, or I don't know, or not. Yes. I think that's how it worked out. I mean, I'd have to go back and bug Mike on his experience. It's probably vastly different from mine because I knew he had a lot of back and forth with them. But we were opportunistically borrowing time from people. We got them excited about the project. They wanted to help. You know, I couldn't put a number on it, but my sense is that it was kind of ad hoc on the fly. Oh, well, here, look at this thing, or let me give you this piece of code that you can run. And then I think it kind of came out in the iteration that Mike did with them over time. Yeah, that makes sense. But my interactions with them were always, oh, yeah, I know. Yeah, that's easy. I know how to do that. And, you know, it may have been easy for them. And it may, in fact, be easy. But I think these patterns still are new. And certainly putting them together end-to-end like this is maybe, you know, what's most novel about the project. Yeah. And so you're saying that the result you got to was it would take basically like 10 seconds to go through this cycle, right? Yeah, and to tweet back out. So if I tweet today as Maya, and I always do it on Maya's account, I don't want my mom asking why I'm looking at dresses. Not that she would mind. Right, right. You would just say like, oh, don't ruin your Mother's Day surprise. Right, right. Sure. But, you know, the results return in, you know, roughly 10 seconds. I don't know how much of that is Twitter lag. I don't know. I've never actually tailed all the logs. I know Mike did that at one point in time. I think, again, what's interesting is that Maya didn't at mention the retailer, right? She didn't say, hey, I'm looking for this. She simply stated, you know, a love for an item and an intent to own it. And we took that positive sentiment and the fact that the item matched something in our inventory and replied to her. Yeah.

So we did an experiment where we tweeted a picture of stereo speakers. And sure enough, we got through and it was like, okay, there's positive sentiment. We should make an offer. Let's go look for an item that matches. Oh, we don't have any matches. We don't sell speakers. We can't do that. So, you know, there are varying ways to come through the decision tree and still not end up with an offer. It wasn't like every single thing went out. But every now and then I'll tweet a picture and I'll get an offer back from our retailer, Boho Chic. That's just random and bizarre. So, you know, it is not a perfect system. But it's something that we, I think, could perfect over time. Yeah, yeah. So, and on that note, like, what are some things that, like, just didn't work out that you're like, well, too bad that that was a failure. Sure. We had grandiose plans in the beginning, right? We wanted to pull in Facebook data. We wanted to have an event on Facebook and a community of people, you know, that interacted with her around that. We thought about doing Pinterest to get a notion of her color palettes or her, you know, deeper insight into her preferences. We thought about, you know, do we build a mobile app? Do we have, you know, all these other ways of engaging her or enriching that data set? And we came back to, no, let's just build an MVP. Let's build the simplest thing that will work, that will prove that we can do something valuable. And that was kind of a dumb stroke of genius in some weird way that we faked a lot of stuff, right? We don't actually have multiple locations to our retailer. We have one, right? We don't have, you know, this rich set of data. We simply have what's in the tweet and in the image. But the fact that we were still able to derive such interesting things out of such a small set, I think, makes it even more valuable. So, you know, I'd love to run an experiment with an actual retailer and actual customers and see what this thing looks like at scale. We've had some conversations with grocers, right?

And we thought, okay, well, how is this? That's exactly what I was going to suggest, right? That's a pretty, like, rich palette of stuff. And then also, I mean, further interrupting you, from a business perspective, my understanding is that the grocery business is all about top-line revenue, right? Because you have razor-thin margins. So the more top-line revenue you have, the more profit you're going to get. So you want to grow volume as much as possible because, you know, you're not going to sell an avocado for $20, right? So you've got to just sell a lot of, like, $0.50 in dollar avocados to make money. Yeah, you may have issues around spoilage. You may have issues or things you want to highlight around prepared foods. I mean, we thought of a couple different scenarios, like soccer mom leaving a day-long tournament. Kids are hungry. I'm tired. Boom. Sweet. Come by, pick up burgers and buns. We got it covered for you tonight. We'll have it waiting for you at the door. Yeah. Or, you know, a guy and his buddies are playing softball. I grew up in Ohio, so this stuff happens, right? And, you know, they just finished the game. They won. They're celebrating, you know, 12-pack on us or whatever, right? Like something that says, come in, grab some stuff. We'll throw something on top just for you. Yeah. It comes from knowing your customer and knowing what they like and meeting them in their moment of need with a compelling offer that makes sense, right? That's tough to do. You're never going to email me something that I'm going to see in time to make that decision. But if you can respond when I'm in that moment, then I think you have a chance of making a more compelling sale. Yeah, yeah. I don't know what it would be like to get bombarded with offers like that. Yeah. Well, I mean, that would be part of what you would have to machine learn it to do, right, is to not be annoying. You need a whole – there probably are like whole data scientists and their job is to make sure is to decrease annoyance. Maybe there's even some fancy scientific name for that of like how do you dial that down so it's not too frequent.

How do you smooth that curve out so that it feels authentic and human without just, oh, offer, offer. I mean, yeah, anyways, not to get into the theoretic weeds. So this is kind of like a jump but kind of laterally related. So especially with the grocery thing, you're reminding me of what was a pretty – what I think was a pretty smart line of thought from the CEO of Jack in the Box of all places. I mean, not because they shouldn't be smart, but I just haven't thought of Jack in the Box in a while. And he was saying that like – I don't know if this is from his quarterly call or whatever, but it was quoted in some article and he was saying like, well, the thing with digital is that once we all get digital, it's not going to be a big deal. And then it's going to go back to like is the product good, right? And are you openly – like the fundamentals of the business. And he's like, you know, and we're Jack in the Box. So sure, we're going to do – I'm adding my own words here. But sure, we're going to do digital. But like we also are Jack in the Box that have this other stuff sorted out. So you should believe in the long-term viability of the company I'm assuming he was getting to. That's an interesting point, right? Yeah, yeah. At the end of the day, it's only so sexy. Yeah, and then so I wonder just to like – to get you to a blue sky solution here or think out loud on it, like what – like given that you – y'all could do this in six weeks, right? And obviously, like let's – you know, we are the increase your cycle time people. But like obviously, it would probably take you at least a year before like it was a thing that was like a major part of the business. That you had it at enough volume and you exposed it and you did the annoyance dial. Yeah, I think that's fair. Yeah, yeah. Yeah, and so like – but you can imagine over a five-year term that like every company could be doing something like this, right? Which again, to the Jack in the Box thing, doesn't belittle it all. But I wonder like how – I don't know. What do you do with that, right? Like – Where does it end up? Yeah, or no, even on a different plane of like how do you plan out for the fast commoditization of something that was incredibly unique at some point, right?

Oh, wow. Like with the grocery store example. Like let's just assume that everyone's going to have this in five years. Sure, sure. Or less. So what do we do? Yeah. Yeah, I mean I have no idea. I hadn't ever taken it that far in my head. You start off with something that's uniquely differentiated and very quickly just gets commoditized. Exactly. Yeah, yeah. And maybe that is the world we live in where those things do in fact happen. I guess, you know, is cloud commoditized? It's probably not that far away. Yeah. Yeah, well, I mean to add a footnote to that, I mean to your point, the things that are – there's a few regular infrastructure things. But the things that, as we've been doing here, that become exciting and differentiating about various public clouds are the actual services they provide, right? Like TensorFlow and is it Marsha or Marcia that Amazon just announced? Like those things are interesting because they're novel new services. Like the idea that, you know, I could start a bunch of VMs is to some extent thrilling, but not that big of a deal anymore. Right. I mean it's interesting you brought up that jack-in-the-box comment because you do come back to what is core to our business? What are the things that matter most to us? Right. And I would argue or assume rather that for jack-in-the-box, it is, you know, the consistent quality experience that you can only get a jack-in-the-box and the predictable time, et cetera, et cetera, right? Like some very core things. And they could certainly maybe, you know, goose the margins on some things or move the numbers in another direction based on a promotion or whatnot. But at the end of the day, their business is providing food at predictable, consumable prices, right? Yeah. And, you know, maybe that's true of retailers and other businesses as well that, you know, all these things are on the margins.

I don't know. We've also seen some pure digital plays. I think of Stitch Fix, right, as really a huge disruptive force in retail. They own no stores. They simply have, you know, warehouses and they have an army of people that help style their customers. And they get that real-time feedback and driven into their automation. So, you know, who knows where this goes? Do we see the end of retail? Do we see the end of dining out? Probably not. I don't know. I'm just babbling now. No, no. That's what I asked for. But I think, you know, to equally babble on my own, right? I think, I don't know. I mean, one of my personal charters is to be the apologist for boring stuff that makes lots of money in what I do, right? Because I don't really know why. It's because there's plenty of people who talk about, like, all the hot, cool stuff, right? And I think it is hot and cool, right? Like, it is the idea of, to stick to retail and groceries, the idea of omni-channel selling, right? Which means there are, instead of just going to the store and buying a product or going to a website and having one mailed to you, there's, like, five or seven different ways that you can mix and match to browse products, find them, pay for them, and then get them, and then return them, right? You can do all sorts of stuff. I once, I forget if it was Gartner, but I remember reading a Gartner report on omni-channel, and they just, they ran the matrix of all the combinations. And it was, it was delightfully overwhelming. I'm sure. Like, if you think about, you know, in-store pickup, order online, order in-store, have mailed you, like, all the different ways you could do it. But, anyways, so there's a lot of interesting innovation to do in that area. And I think even the terms we use to describe sustaining innovation are kind of judgy, right? Like, this idea that just doing incremental or sustaining innovation is all you're doing, like, certainly isn't as cool as, like, you know, robots or self-driving cars.

But it is, like, I mean, obviously you have to do both. And I think it's equally interesting to figure out, like, so going back to the jack-in-the-box thing, right? Like, what are, this loop that y'all have, like, how would you apply that internally to, like, optimize how they're running things, despite our top-line revenue comment earlier? But, you know, there's, it seems like there's a whole lot of other things that you could apply it to, which would grow the business value that you would have. I mean, and there's simple things like, you know, here's a new type of burger you could sell or something. But there probably are all sorts of other interesting applications of a loop like this. And the same benefits that you have of that, it would take you six weeks or less to just try something out because all of this is fit together. I don't know. It gives you an interesting sort of tool to start doing incremental innovation or sustaining innovation. And I think if you approach it in those terms of let's run some experiments, let's see what we can learn. You know, to me, that is the most compelling thing. I came into this thing relatively cold. I mean, my wife has done fashion retail for many years, studied fashion design in college, grad school, I should say. And so that is her world. And I grew up in the test bed of retail in Columbus, Ohio. So I'm certainly not unexposed to it. But, you know, with just our bare minimum of knowledge, we were still able to create something compelling. I think about experts, domain experts, or people with the business experience and, you know, retail or grocery, et cetera. With that domain expertise to be able to put these things together in ways that benefit them, that's the most compelling piece to me. Yeah, no, for sure. For sure. So I'm a novice and we came up with something cool. What could an expert do with it, right? Exactly. Get someone who's been in the field for a long time. Well, that's great. Is there anything else you want to add on before we wrap it up? I can't think of anything, Coté.

I hit my main points. Had a delightful conversation with you, as always. Yeah. This is really fun. I don't know how you get to do this day in and day out, but what a cool job. Well, you know. Do you have any metaphors to share before we go? I'm all at it. Have you exhausted? This is about 3 p.m. on Friday. It's like I have a deep well, but it's been a long week. And you're on vacation next week, correct? Oh, I'm traveling. Traveling. Okay, got it. My wife will think it's a vacation when I get to go to bed in a comfy hotel bed without kids. Have you ever taken her on one of those trips? Almost. Almost. Okay. We're getting there. That changed my wife's opinion of my business travel. Yeah. So, yes. Not as glamorous as it seems from the outside. Yeah. But still pretty cool. She once went on a – and this doesn't help the case that you're trying to build. She once went to Las Vegas with me. Right, right. Which is not quite the slumming it you would be looking for. True, indeed. Yeah. But so if people want to follow up with you, you got a Twitter handle or use the CB radio or anything? I got a at SCT, my initials, and I'd be delighted to continue the conversation there. You're going to share some materials, I think. Yeah, yeah. You'll have to send me some links of other stuff. I've done a video and a demo of all the other stuff. And you certainly have the original tweet that spawned this conversation. Yeah. Yeah, you're like me. You must have gotten into Twitter early to get a three-letter thing. Yeah. I mean, I think I was in it even earlier, and then I quit because I was like, who's ever going to tweet? This is weird. Yeah. And then I came back shortly thereafter, but still it was available. Yeah. Yeah. Yeah. You know, it happened to line up with my work when I was at Red Monk, so I kept using it. But there was this point where I was like, all right, I don't need to be getting – every time that Chris Messina thinks of something, I don't need to get a text message about it. Which was way back in the – what was it, 4-0-4-0-4 days? I remember. Yeah, the farewell. Yeah. Those are good times.

But yeah. Absolutely. Well, great. Well, thanks for being on. And with that, this has been another episode of Pivotal Conversations. If you want to get the most recent episodes of it and kind of browse our back catalog, we've got a – to recommend some episodes, we've got now a good collection of what I call, I don't know, Pivotal IRL. Like people like Allstate that we've spoken with and other people who actually are the end customers, not just theorizers like ourselves here. Applied theorists. But, you know, there's various collections there that you can go get. But if you go to SoundCloud.com slash Pivotal Conversations, you can find those episodes and the RSS feed to subscribe to. And more or less every Thursday, I go and post an episode formally, the show notes, over at Pivotal.io slash blog. And I've been putting transcripts up there as well if you're interested in that. And also, we've got a couple of other podcasts you should check out. We have one called That Moment, which goes over people kind of in a more – in a different style than we've been talking about here, kind of discovering when they should do something different and kind of walking through how they've been doing that. And then, of course, there's the Pivotal Insights podcast, which sometimes for a while nominally covers sort of the data science and things like that. But they also cross over into just general-to-purpose things like the most recent one, which I mentioned here before that I like, talks with one of the people who heads up our application transformation and modernization re-platforming service. And there's a really good – in the same way that we've tried to explain a bunch of visuals here – there's a good explanation of how you methodically go through the portfolio and decide what to re-platform. So with that, we'll see everyone next time. Bye-bye.

Tanzu Talk · 🍕 Get more pizza.