Today on How AI Happens we are joined by the leader of digital change at UPS, Sunzay Passari. You’ll hear all about Sunzay’s extensive career, how he landed his role at UPS, how he plans on driving technological transformation, and how AI has made its way into the company in recent years.
Our guest goes on to share the different kinds of research they use for machine learning development before explaining why he is more conservative when it comes to driving generative AI use cases. He even shares some examples of generative use cases he feels are worthwhile. We hear about how these changes will benefit all UPS customers and how they avoid sharing private and non-compliant information with chatbots. Finally, Sunzay shares some advice for anyone wanting to become a leader in the tech world.
Key Points From This Episode:
Quotes:
“There’s a lot of complexities in the kind of global operations we are running on a day-to-day basis [at UPS].” — Sunzay Passari [0:04:35]
“There is no magic wand – so it becomes very important for us to better our resources at the right time in the right initiative.” — Sunzay Passari [0:09:15]
“Keep learning on a daily basis, keep experimenting and learning, and don’t be afraid of the failures.” — Sunzay Passari [0:22:48]
Links Mentioned in Today’s Episode:
Sunzay Passari 0:00
Trying, and don't be shy of some failures, and I'm sure you'll be there. It's not no rocket science. It's just a word of persistence.
Rob Stevenson 0:08
It's not rocket science. It's just AI.
Rob Stevenson 0:15
Welcome to how AI happens, a podcast where experts explain their work at the cutting edge of artificial intelligence. You'll hear from AI researchers, data scientists and machine learning engineers as they get technical about the most exciting developments in their field and the challenges they're facing along the way. I'm your host, Rob Stevenson, and we're about to learn how AI happens. All right, everyone. Welcome back to how AI happens. The only AI podcast that doesn't insult its listenership by pausing to define Lidar and drift and llms and all those other things that those podcasts are doing to you. We assume you have a base level of literacy year on how AI happens, and it's a good job we do, because today's guest is a technical one. He's had tons of experience in our space. He was recently the Global Head of project management for eBay. Currently, he serves as a Senior Director and Head of innovation and transformation over at UPS. Sunzay Passari. Sunzay, welcome to the show. How are you today?
Sunzay Passari 1:20
Thanks, Rob, thanks for having me doing great. Nice to talk to you.
Rob Stevenson 1:23
Yeah, yeah, you as well. Did I do your curriculum vitae justice? There is there any more about your background that we ought to know before we jump in here?
Sunzay Passari 1:30
No, I think you did a right job. And, yeah, I'll love to talk a little bit more as and when you give me a chance.
Rob Stevenson 1:36
Yeah, definitely. So how did you kind of wind up at this role here at UPS
Sunzay Passari 1:41
pretty interesting actually. You know, I've been in this journey of driving innovation or quarter century, almost 25 years. I mean, I started way back in 1990 and I got an opportunity to work across various companies and largely around the telecom and media space, but working along with other companies in all kind of verticals to drive innovation, right from the days of dotcom days boom, if you remember, and then there was a mobility and there was a cloud, and I've seen this all journey. So it has been a passion for me to be working in this space over last 25 years. And few things happened, like there were quite a few first industry first I got a chance to deliver and launch around the world. And second thing was, I got a chance to work around the globe, which includes North America, which includes Europe, which includes Middle East, which includes Asia, large part of Asia, in India, like India and China, big geographies. And last but not the least, even in Africa. So that was kind of my background. And ups. Was looking for someone with a very, very external perspective, someone who can actually bring so UPS a little, little about a UPS as a as a company, which is 116 year old. So there's a lot of legacy, and, you know, lot of legacy. They carry both lot in terms of technology, in terms of people, in terms of processes, in terms of culture. So they were looking at someone who brings a real external perspective from all perspective, like, you know, technology or people and processes. And for me, this was a great company, because with that kind of legacy and that kind of a history, just you can imagine the complexity and the depth it brings to the whole initiative, because we exist in 200 plus countries, and we are like employing 650,000 of people, 150,000 vehicles flying across the globe, 24/7 so just you can imagine the complexity and and that complexity is the right use case for driving AI, or edge computing, cloud computing, etc. So yeah, I just jumped into it. I said that's the right platform to do something bigger and better, you know, in near future. So that's how I landed here.
Rob Stevenson 3:50
This is why I was excited to talk to you, Sanjay. Because when you hear 116 year old company, when you speak about the scale, like you just did in over 200 companies, hundreds of vehicles, etc. Don't think innovation. Frankly, you don't think like agile. But yeah, here you are saying that this was the perfect sort of opportunity to be driving some AIS and transformational change. So why is that the case that you find that this huge, this huge scale, this huge operation, is the right opportunity to drive some AI and transformational change?
Sunzay Passari 4:19
Yeah? Because, as I said, the complexity brings the opportunity here, the complexity of our entire operations, the complexity of our skill, the complexity of unknown variables which we are dealing with day in, day out, whether it is a weather condition or weather availability of right kind of resources, whether it could be some political conditions. Recently you saw that, you know, there was a war between Ukraine and Russia. There is an industry in the middle. So there is lot of complexities in the kind of global operations we are running on a day to day basis. For reference, like we deliver 26 million packages every day. You know, carry 26 million so that that complexity brings in huge. Huge opportunity for AI use cases, generative AI use cases. And that scale and complexity both actually, that's what I would say.
Rob Stevenson 5:07
Yeah, it makes sense. The more variables at play, and there are so many. And as you say, from just the actual human workers to these geopolitical happenstances. And so the more variables at play,
Sunzay Passari 5:18
including weather conditions, which are unpredictable
Rob Stevenson 5:20
Right, of course. And this is, you know, I guess, as you say, ripe for a machine to kind of Disrupt. What does that look like over the last several years? What is kind of the AI disruption? How is it taking form of UPS?
Sunzay Passari 5:18
Okay, it's slow and gradual in the sense that, let's, I like to try to break AI into two pieces, AI and Gen AI. Now AI one we have been one of the pioneers and and the route optimization game, especially, you know, if you have a heard our famous we are famous for our drivers, for one thing, like our drivers never take a left turn. You know, by design, they're supposed to take a right turn. Now, this is one piece of our route optimization. We are we've been using massive AI in terms of gathering data, gathering traffic data, gathering weather data, gathering load data, whole, whole bunch of data which is being optimally. You know, used to optimize this so, so these kind of use cases. And also now think of our operational optimization in which is going in terms of our sorting centers or warehouses. And again, the lot of pieces moving. As I said, we are moving like 27 million pieces every day around the globe. So every warehouse is seeing like 1000s and 1000s of pieces every second we moved from one point to another being sorting. So this piece we have been doing for a while now, and we are one of the pioneers and leaders in this now cut back to 2023 where the Arc Light shifted to a generative AI use case. So now imagine the scale of operations, the amount of data which we produce on a daily basis. Now put just marry this together, and this creates a perfect use case for you. You know, perfect use case for exploring generative AI. Now there are few areas which we are doing. One of the lowest hanging fruit we talk about is in terms of a customer enhancing our customer experience. You know it can craft personalized communications for our customers. Imagine a chat bot that can not only answer your basic questions, but generate a tailored shipping, even a tailored shipping course, or a proactive update on delivery delays or weather events, Route Optimization, I have talked about it. Then there is a proactive resolutions of of issues and which, which is both on the B to B and B to C side. It is a problem with deliveries or missed packages, or damaged packages or or even when we work with lot of our partners, they are like machine is talking to machine. So we have to preemptively identify whether there is a likely connection breakage or any kind of a, you know, API failures, etc, so that we can resolve them. And with the size of people I was talking about like, see, we are employing 650,000 people. There is lot of training requirement, and there is a lot of content requirement for our workforce itself. So we are using generative AI to create a very personalized training material, training content for them, which we can do at their own speed, at their own complexity. So there are a whole bunch of them. So while I speak today, and we are, I would say at a very, very native stage of exploring Gen AI. And these are just few use cases, I'm sure as we progress, we'll have far more like one thing which I'm personally driving is about, you know, insights, insights for customers, insight for operations, where a leadership, or all the managers, the VPs and the directors, can just go and kind of talk in a, you know, natural language. They can talk to the system and get all those insights which are very crucial for our operations. I can keep talking. There's a long list here. Rob,
Rob Stevenson 8:45
yeah, we'll pause there quickly so I can chime in. But that you mean you were about to, I think, answer the question I wanted to ask, which is, when you have a company that this of the scale, where there is so much opportunity to inject advanced tech like this, AI, is often this, like magic wand, that various leaders throughout the business think you can wave over their problems and it will just fix them. And they may be right in some cases, but when you consider the scale of UPS, how do you decide, okay, this is the area where it will be the most disruptive. Here's where we're going to make this attempt to implement new tech.
Sunzay Passari 9:19
Very good question. Ram, there's no magic wand, okay, people like us who are actually delivering on the ground, there is no magic wand as such. So it becomes very important for us to bet our resources at the right time in the right initiative. We just can't be riding on a hype to bet our resources or talk about generative AI and to talk about its potential in the boardrooms, it's great, but when it comes to actually, so we are very data driven, as I said, you know? So there are two kinds think of two ways of looking at at our initiative. One is a qualitative part and one is a quantitative part. So while we are investing our resources, the qualitative parts, you know, some of the areas we touch upon are. How much of it is future proofing our legacy? How will it optimize our network? Because that is one of the largest, largest element in our entire operations. How is it going to personalize? Is it empowering our workforce? Is it kind of initiative which will drive sustainability and efficiency? Is it secured enough so these are on the qualitative sides and on the quantitative side. Of course, we are pretty conscious about every penny we spend, so we like to do a cost benefit analysis, and quite in depth actually, you know, we go through that, and we are not only do a cost benefit analysis before even starting an initiative, but we assign a number to each of this initiative in terms of either generating revenue or bringing the cost down, and that is monitored over next three to five years of that initiative. So it's a pretty scientific approach to how we deliver an initiative. Take an initiative and drive it actually.
Rob Stevenson 10:54
Could you share a time where you maybe decided, hey, this area of the business is not ripe for for this kind of tech where it's like, we're not gonna roll out generative use case in this scenario,
Sunzay Passari 11:05
a lot of them, actually, my answer is more often, no, it will be hard for me to pick really specific things, like, there's a lot of hype and in the boardrooms also, you know, with this all media hype, things are going around, And then you get a ping from your executive leadership. And hey, you can we explore this? Can we do generative? Yes, we can. But let us figure out two things here. A What does it takes to do and what does it actually deliver? What does that benefit? It's delivered to the company. And like, if you are good on both the parameters, like, what does it take to do it? It doesn't take a mountain to move a mountain to deliver that, because I'm personally being really conservative when it comes to driving Gen AI cases, use cases. So rather than saying what I'm not doing, it will be more appropriate. It will more easier for me to say, what are the areas I'm focusing upon in terms of Gen AI, because what I'm not doing is far too many of them, and I'm saying no to it.
Rob Stevenson 12:02
Yeah, this episode of the podcast is called how AI doesn't happen. Yeah? But I'm pleased to say that because like, like, you say you're being pulled in lots of directions. Executive leadership says, let's do some generative and you're like, that doesn't mean anything like, and it seems this idea of, okay, what is it going to take to pull that off? And what is the benefit that maybe isn't a technical answer, is that more of just like, oh, this means this much money spent on Compute and this many more hires for this level of reduction in costs, or it seems like it maybe isn't an AI expert answer, is that, right?
Sunzay Passari 12:35
I mean, I would say yes and no both, because AI rides on an engineering backbone. Now, what is my readiness of doing a specific kind of use case? That is the first question. So I mentioned at a very high level, what does it take now to evaluate what does it take? What is my AI readiness? Now, AI needs lot of data, or generative AI leads far more data to train your llms, and, you know, get a meaningful output out of your LMS. You're not doing it just for the E codes. It's not a lab project for us. No, no. Project is a lab. It's a right live, commercial project for me. Now that is the even most of the requests get failed in the first level of filtering. If I use that word that do I have right amount of data or right kind of data and right plumbing of data required to run even a SLM, if I have to use that term, or an LLM and train an element, or, you know, use, or kind of, you know, deploy and rag inside my infrastructure to get a meaningful output. So today, as we speak, I think my personal focus is more on getting that data engineering piece rock solid, because once that base layer is rock solid. You have right data, you have data integrity, you have data security, authenticity, everything. And then then you have right slices of data, which is required you have created, let's say data marts for different use cases or different functionals. Now they are talking to each other. That's a larger piece. And trust me, Rob, and it's not just I'm saying on behalf of UPS, most of my peers, counterparts in large companies, are dealing with the same issue. So while we are not shying away from deploying small, little use case, experimental use cases for generative AI in these areas, as I was talking about, deriving inside customer insights from my customer data, that's a one small use case, but in parallel, the larger piece we are working is facing the data engineering piece. So that's where the key focus is. If that makes sense, that if that's you find is a technical answer to you,
Rob Stevenson 14:30
yeah, yeah, certainly. So is that why you are conservative when it comes to generative use cases?
Sunzay Passari 14:37
Yes, I am. Yes. I am, because it's very easy to get swayed away in this whole hive, and that is one of my key roles, to say no wherever it makes sense, because saying yes is an easy part, but saying no the hard part, because when your leadership says they want something and and you have an audacity to say no, then you have better data to prove that why you are saying no. So that's hard part. Actually. You.
Rob Stevenson 15:00
Yeah, yeah, of course. Well, I'm pleased to hear that you're saying no to lots of these things, because, like you say, there's no magic wand, but you're not saying no to everything. So I guess you know, we might as well start with generative what are some of the generative use cases you you think are worthwhile?
Sunzay Passari 15:14
Two key for me, I can, I can talk about for us, actually, two key use cases. One, we are deploying in our in our customer interactions? So we have very small segment of retail customer interaction. Most of our businesses actually B to B, so we may not need directly interact with customers, except for a very small thing, like, you know, where is my package or when it will get delivered. But having said that, even that small is large enough, you know, because considering our scale, that's largely different. Becomes more complex if you're talking of many different geographies and many different languages and many different cultures and many different compliances and laws and all of that. So China is so very different from what's happened in Europe, etc. So we are actually deploying some of these in our customer interactions customer care systems to loosely term them. Second thing, as I was mentioning, is about deriving insights from customer data, what kind of volumes and what kind of planes they are using, what kind of pricing, what is the price elasticity and discovering about you know, what are those parameters where we need to play around which will help customer win as well as us win which help us grow volumes with the customer. Customer feels that he is growing value when he's working with the EPS. So this is one of my key area. And also, while we are doing this price discovery, price pricing is a very complex subject in a logistics especially in our kind of company where your scale is too big, your network is too big, you know, and there is no one size fits all. It's like, more like a surgical pricing today. Now for us, that is, for me personally, now is that is one of the key initiative to discover that, you know, what's the stretch price to right price point? Because you'll be surprised, and most of the common people don't know, apart from a base level price, which, you see, we have something called 300 exosorials. You know, that means there are 300 different charges which are potentially leave it upon a customer, which could be levied upon a customer. Now, just imagine the complexity of just managing that. And not to talk about those infinite number of cells of pricing, you know, this is the weight, or this is the dimension, and this is the zone. We call it zoning. And so this way, this doesn't mention this zone. That's the price. It becomes really complex. So simplifying and discovering the right price is one of the use cases we are trying to try, you know.
Rob Stevenson 17:34
And so that would be for UPS, is like data scientists querying your own data, because the alternative is exporting data to a spreadsheet, or knowing are knowing some other kind of language, like, Why do you think it'll make it easier for them
Sunzay Passari 17:47
For for data scientist, or for the end user, assuming to so for, for my business users?
Rob Stevenson 17:51
Well, in the case where you're trying to pull insights out of customer data, I assume that's for the benefit of of UPS people, like internal right? Yes.
Sunzay Passari 17:59
So we have customer data, what kind of volume? So we might have historical data, or we might not have a data about a customer because it's a new data. Now, we are applying lot of data science like, let's say, AI models, you know, learning, machine learning, models on top of that data, to figure out what makes sense in two ways, like, what makes sense in terms of those price points for a given service, for a given parameters, let us say, I mean, there are many parameters. I am trying to make it very simple for given parameters, what is that absolute price point? Which makes sense? And then to grow that know, what is that deviation, or what is that elasticity in that and second thing is the simplicity, because the way logistic industry has evolved over last 100 years. Actually, the pricing has become far too complex for a common customer to understand that. And when I say common customer, it includes customers like Amazon or a Walmart or a Costco, which are like themselves, very big companies, but even for their trained managers, it becomes hugely complex. So that simplicity will actually drive a lot of adoption or a lot of volumes, I would say. So we are working on both fronts.
Rob Stevenson 19:06
Yeah, I'm glad you called out the B to C versus B to B difference, because for myself to walk two blocks to my local UPS office and ship a package, I can do that, and I know that it will work and it will arrive, right? I assume that's not where UPS is bread is really buttered, right? Like, you have much, much bigger customers than me, and so is that the priority, like these bigger kind of corporate customers will receive some of this tech first. Will it ever dribble down to me? Will I ever be able to have a conversation with an LLM about where my package is at?
Sunzay Passari 19:36
Yes, you will. And that's the first thing, which I told you when I was talking about, you know, a customer interaction that is a low hanging fruit. Most of the companies are doing that. So are we in terms of how to make that experience very seamless for you. Because even if you from a customer perspective, you are a small customer. There are two things. Your package is the most important thing in the world to you. We want to make sure that that is recorded as. Much priority as any anything else, because all these B to B customers are eventually serving people like you and me. So I talked about Amazon or a Costco end of the day, they're delivering it for you. So we keep our end customer in mind to whatever we do. So even if, while from an implementation perspective, it might be that, you know, a Walmart or a Costco or an Amazon is implementing that, but eventually the beneficiary is the end customer, which is you, which is a common man on the street to make sure, and second thing to specifically to answer. So you walk into an UPS Store. That is a segment for us by itself, and that is exactly where the customer interaction is happening. There is the only touch point, I wouldn't say only the primary touch point for most of the customers to know what UPS is, but everything else is in the back end is much larger. So we want to make ensure that that touch point is well served. And you, you walk in and walk out with a greater experience, nothing less than wall,
Rob Stevenson 20:54
right, right? Of course, there is this curiosity, though, and like when you are speaking with a chat bot, or it's kind of in service by GPT, a little bit where you can query it for data that you maybe shouldn't have because it's based on what other people have uploaded. So there has to be this level of like of privacy or tiered access to information. I assume that that's similar, because, you know my experience right now where there's like a website I can like, in certain there's all these various like updates on where my package at is that I assume there's far more data on my experience on my package, but there's that challenge that, like, if I had free access to ask anything I want, that I could just see it, but yeah, I assume you wouldn't want that compliance. Sure. So how do you go about putting that into a chat bot so that I can't get all these this non compliant data from the chat bot?
Sunzay Passari 21:39
So as I said, there is whole lot of masking of data, which happens even before it is exposed to any LLM, I mean, for a company, again, like our size and our complexity, that is our key, our shops really start if we don't take care of privacy and secrecy of the data and completely follow PII regulations in every country, not just us. So if you are in Europe, in China, India, wherever, even in Africa, every country has their own privacy laws, and we have to comply with that. So the data as I if you remember, I was talking the data engineering piece. So this is one of the key things for us in terms of data engineering, that if we are exposing any data to our llms or SlMs or rag architecture, whatever it is, we make sure that none of PII is exposed to any of these technologies so that so we kill it at the source.
Rob Stevenson 22:29
Actually, I love that. It's such a simple answer. The answer is, don't train the Chatbot on it, right? Yeah, exactly, and then there's no risk of it getting through. Yeah, exactly, exactly. Yeah. Makes sense. Well, sunset. Before I let you go here, I just wanted to ask you to maybe pass on some some sage wisdom to the folks out there listening for people who are wanting to find themselves in a position like you at a large company, making decisions at a huge scale on how technology is rolled out. What advice would you give them as they look to forge their career?
Sunzay Passari 22:57
Pretty interesting. I think learning. Keep learning on a daily basis. Keep experimenting and learning, and don't be afraid of the failures in whatever role you are, even if you are a student. I would, I would strongly encourage that this, AI, is the fourth dimension. I would say, you know, there was a compute dimension, which started in 60s, like, you know, if you see first 30 years, 60s to 90s, there was compute, and then there came connectivity, when Internet came. And then came ubiquity, along with that, which was mobility, actually. So during that 30 years, the next 30 years, three things happened. So there was a compute, there was a connectivity. The connectivity came with ubiquity, and then emerged the cloud. Okay, now the fourth dimension is AI. This next 30 years, the world will be technically changing beyond anyone's imagined imagination. Today. Even if I'm leading a large company's innovation, I can at the best guess what the technology would look like from five years from now, but I cannot predict for sure. So keep learning on a daily basis, keep trying, and don't be shy of some failures, and I'm sure you'll be there. It's not no rocket science. It's just a word of persistence.
Rob Stevenson 24:06
It's not rocket science. It's just AI,
Sunzay Passari 24:11
exactly. I think that these terminologies will change over a period of time.
Rob Stevenson 24:14
Sunzay, this has been fun. Thanks for being on the show and for sharing me all about your position. I've loved chatting with you today.
Sunzay Passari 24:19
Thanks, Rob, thanks for giving me this opportunity, and I hope your listeners like the talk, and I'll be happy to jump on anytime in future if you want me again. Thank you so much.
Rob Stevenson 24:29
Don't threaten me with a good time. Sunzay
Sunzay Passari 24:31
you too. Bye, Now
Rob Stevenson 24:34
how AI happens is brought to you by sama. Sama provides accurate data for ambitious AI specializing in image, video and sensor data annotation and validation for machine learning algorithms in industries such as transportation, retail e commerce, media, medtech, robotics and agriculture. For more information, head to Sama.com