Founding a Company in the AI and Analytics Space

 

David Hailey, founder and CEO of Countifi, started the company after seeing a need for more efficient inventory management systems during his time as a general manager at Delta Airlines and as a manager at Ernst & Young. Countifi uses image analysis to manage inventory rather than manually counting items. The company’s tools are currently being used by hospitals, airlines and universities, including Delta Airlines and Duke Hospital.

Hailey, Weekend Executive MBA ’13, explained to Professor Jeremy Petranka, how he founded a data science-based company, despite not being a data scientist himself. Hailey also discussed his application of computer vision to streamline inventory management.

The following excerpts from the interview are edited for length and clarity.

JEREMY: Countifi—the first time you talked about it, it really did make me just think, "Ah, that is a perfect use case." It's a perfect example of how transformational, fairly off-the-shelf data science techniques can be when they're applied to the right business needs. I also think that your journey personally is really inspirational for people who kind of see the needs the way you did, but don't necessarily have a strong data science background and are worried that they can't make this type of pivot. So, with that though, let's just get started on Countifi itself. Can you give everyone an idea of what the problem Countifi was designed to solve and how it solves it?

 

DAVID: Yeah, so during my career, I've spent, as you said, time in EY, time in Delta—did a lot of manual inventory counting that usually always sucks. They are always at the wrong time. Any former Big Four accountants out there know that you usually can't plan New Year's because on January 1st companies have their 12/31 year-end, you have to be somewhere counting something at some remote location. And so, doing that for a number of years and then realizing while I was in the airline industry that the entire industry—there's a lot of improvements that can be made from a counting perspective.

 

Your big, large airlines can spend close to a billion dollars for the type of inventory they put on the plane. That's the food, liquor, beer, wine that you enjoy while you're on the plane, but [airlines] still don't have a good sense of controls over those items. And if they do, a lot of those controls are manual, right? Someone is somewhere manually counting those items—one, two, three, four, five—and it just takes a lot of time. If you think about manually counting 2,000 miniature liquor bottles, right? You'll probably be drinking after you count like 100, right? Because it's just so intensive.

 

So, I thought about with that problem, especially with the advances with technology, how can we form something or solve that problem and use technology to do it? And that's how computer vision came into play, just because advances in computer vision. Everyone has a smartphone in their hand, in their pocket now. Advances of other technologies that we use and can use that to solve that problem, help bring those manual processes, add some tech to those manual processes. Make it easier, simpler, and save time.

 

J: So, what's the idea behind Countifi? Are you basically using image analysis to look at the items coming through as opposed to a person counting them?

 

D: Exactly, so we try to not change behavior too much. For the airline example, you think about those beverage carts that come off the plane. Instead of what usually happens, they come off the plane, they go back to a big kitchen facility and this happens almost at every airline all around the world. They come back to a big kitchen facility and they get them ready for the very next plane, right?

 

So, if you think about a beverage cart that might have five different types of sodas whether it's Sprite, ginger ale, Coke, Coke Zero, and Diet Coke. Those five—somebody will look and say okay, how many do we have left and how many do we need to go to the next plane?

 

What our process is, is we have a camera inside of that kitchen facility, and as those items come in, before that person puts or refills that tray, they either hit a button or it does it automatically, it'll take a picture of that square. Takes a picture of that square and it counts the items, and they just move on. It takes about a second. So, it doesn't add any additional real-time to the overall throughput.

 

The data itself is usually not needed by that person, but somebody in supply chain. Somebody in finance will be the one using that data, so we have a dashboard that they'll look at and say, “hey, the flight from New York City—from LaGuardia to Atlanta—we see that 75% of the Bacardi rum or the Coke Zero is not being consumed. So, what do we do?”

 

We give them more data to allow for better decisions. So, if those items are not being consumed, is there a way that we smart or optimize our inventory? And a lot of times we see that, and this is, like I said, almost every airline you see. They're over-ordering because they want to make sure the customer experience is optimal. But if there's a way that you can smart order—order based on actual data—we'll give them data back that says—okay, this Bacardi rum or this Jack Daniels, 65% is not being consumed or even 30% is not being consumed. Is there a way that we can reduce this inventory? Or we see that the Grey Goose is being consumed at 100% and we always run out, right? Is there a way we can pull levers to make sure that one, the customer’s satisfied—and two, we right-size our inventory which usually ends up in some type of inventory reduction?

 

J: And it feels like again, just kind of thinking through that process that any inventory process that has high turnover and fairly standardized products with a process that goes through a physical point, it feels like this is just perfectly aligned with anything in that ballpark.

 

D: Yep, it is, it is and you think about even the people that use [radio frequency identification] RFID. RFID's a big thing now that a lot of companies will use and we did a lot of research with RFID, but the issue is from that perspective, and that same kind of environment you just explained, has one access point in and out, and is there an easier way to count?

 

But what we realized is even with RFID, you have to have a tag on each individual item, right? So, you think about your big airlines might go through 100,00 miniature bottles, 100,000 cans of Coke per day, right? Because of that, that is where the limitation of RFID comes into play and that's why we use computer vision to help solve that problem. But you're right, it doesn't have to be airline items.

 

We have health care clients as well where we will put a camera into a store room and instead of a nurse, or somebody in supply chain, or somebody in inventory management coming in and manually counting those items, we actually take the human element totally out of it and we'll install a camera inside of a store room. We’ll take pictures several times a day, three or four times a day, and then without going to the actual storeroom somebody anywhere actually around the world can look at that dashboard and know exactly how much inventory is in that room.

 

So, it's really taking the manual process and adding technology to it to increase efficiency. Especially when you have health care space, we see a lot of labor shortages, a lot of labor issues as well. So, we give that time back so they can do something more efficient like save lives or everything the hospitals do.

 

J: I want to talk about how you got there because I think that was not surprising, but so impressive that you're the founder, you're CEO of primarily a data science-based company. That's the core technology, but you don't have a traditional data science background. Can you walk me through how you recognize that there's a problem that needed solving, but also how you realize that data science could get you there if you weren't actually aware of all the things that data science can do?

 

D: Yeah, that's an interesting question. So yes, I'm not a data scientist, not a computer scientist. Do not code at all. Do not code any of the languages, but I knew from a just overall computer vision perspective that the advancements in computer vision for the last 10 years have been immense if you think about it, right? You can open the Amazon app now and go to the grocery store and take a look at a barcode. It'll tell you exactly where those items are and what the item is, what the price is, how long it’ll take you to get there. So, computer vision is leaps and bounds where it was even five years ago and everybody has a smartphone in hand now, right? (Everyone) has at least a 10-megapixel camera which you get can very clear, very good data.

 

In my case specifically, I knew what the problem was. I understood the problem, understood kind of the overall value proposition, but you're right, I did not know how to solve it. There has to be a way, so we looked at RFID. I was talking about RFID a little while ago. Is there a way to implement RFID to solve this problem? We even looked at, from an airline perspective, we even looked at robot arms. Is there a way that we can use a robot arm to move these items? But a lot of those are just too capital-intensive.

 

I was actually in an entrepreneur group and this is before I even had ever started the company, and I was kind of talking about the problem I was trying to solve and actually one of the professors, I think it was a Georgia Tech professor, mentioned computer vision. And so, as we started talking about computer vision capabilities, I was like, “Oh, that makes sense. I think we have something.”

 

And so that's where it all started. It was really putting myself in an environment of other entrepreneurs, other coders, so to speak, and that gave me the idea of, or someone actually told me that you can use computer vision. That kind of gave me the idea to start researching more.

 

J: One of the things that I wanted to ask you is [about the] decision to start bringing on data scientists. There are kind of different views on whether—I don't want to call you a data science company because the actual inventory, rethinking the process is kind of the core offering with the data really enabling that to happen. That said, the core underlying technology is data science, and [there are] kind of different thoughts on whether the founding team has to have a data scientist on it in terms of a co-founder, or whether you can actually just hire them as employees. You've navigated one version of that and so you have some kind of, the benefit of hindsight. What are your thoughts on that decision?

 

D: I mean, I think just because you have a data scientist—a core data scientist—doesn't mean you have a successful company, right? And so, I think if I would've known now what I know before I would've potentially brought on a data scientist earlier, but they're not cheap. So, our lead data scientist has a Ph.D. in computer science. And [there are] not that many of them in the country, and a lot of them know their value. I think part of that is—I don't think it's necessary to have a data scientist as a co-founder, so to speak, but they need to be pretty close. They need to be pretty close, especially if you are a database company. Their skillset is what you need to pull approved clients that you can deliver value.

 

I think when we look at it, it's two things. When we look at it—it’s kind of sales and awareness, especially from a startup perspective, how many people or how many companies can make sure they're aware of what we do, right? And then how can we make sure that we have the knowledge and the skills, and really the proof, to back it up—that we can actually do what we say we do, right? Because you’ll have a lot of AI companies that are not really AI, right? There's a team of people somewhere in some closed or remote place that are really just doing a bunch of manual counting. That's very, very cheap.

 

So, I think to answer your question, you don't have to have a data scientist—in my opinion—as one of the co-founders, but they need to be very close. You need to make sure you can bring sales in the door because without sales, without revenue, without that client awareness, you're just a bunch of people in a room just talking about stuff.

 

J: One of the things that I think is idiosyncratic about AI and machine learning in the startup world is that creating a minimally viable product (MVP) can be tricky. Normally we think of MVPs as getting something out there so you can start testing your theories on actually what customers want—what they're willing to buy, what works, what doesn't—as you start adjusting your strategy. But with pretty much data science inherently, you have to be able to train it on data that makes it good enough. And the startup might not have access to data. You might also not have the resources to do a fully fleshed-out production model. So, as you were in this process of finding initial customers but getting it to market, how did you balance the need to get to market, but also having your data science on what you knew was solid ground which again, can have this longer runway?

 

D: We take more of a consultant approach to what we do. Part of that is understanding the current model—not the current computer vision model, but the current process of what they're currently doing. How are you currently managing your inventory? Are you manually counting these items all the time, right? And so, even if you don't have a model that's ready today, we know that we can add some value over your current process.

 

We just talked to a company, a potential client, yesterday that does physical inventory. They have a counter and a recorder, right? That counter and the recorder goes around the warehouse and counts those items. If they get anything wrong, then a different counter and recorder have to go out and count and look at the items that were missed. Then, a fifth person has to go out and reconcile those two teams. And this is a very sophisticated company.

 

So, in that one process—and they do this monthly—there's five people that have to count those items. Even if we say—hey, we might not have a model that's ready tomorrow, we know that we can start with the process and get better than what you're currently doing. We can send one person out. Tomorrow, we can start with our mobile app and send one person out and start doing those counts. Then we can start building that model, start gaining our data sets to help count.

 

And the good part about what we do is once you take those pictures, you still have all the metadata associated with those images as well, so you can look at them anytime. I can see that David, on David's device, he took a picture. It was April 1st at 3:30, right? I have all that data and it's always automated. So, you can always go back and verify.

 

To answer your question, kind of being very transparent that you would not have an overall computer vision model today. But compared to your current process, we can make sure that we're more efficient, giving value, more accurate, and cheaper than your current process.