How AI is Rebuilding Recycling- And Changing the Economics Behind it

Monisha Saldanha (00:00) Welcome to Colorado Tech People, the podcast where we talk with the founders and innovators building companies that shape how we live, work, and connect. I am your host, Monisha Saldana, an executive with 15 years of experience in product management. Today's episode explores how technology is transforming recycling. I'm joined by Joe Custard-Garrett. my gosh. Joe, how do I say your last name? Joe Castagneri (00:27) Castanerri, it's like lasagna. Monisha Saldanha (00:30) Castanerri. Did I just say it right? Okay, I'm gonna do it. I'm gonna just start from the beginning and I'll just edit out what I said. Joe Castagneri (00:31) Yeah, yeah, yeah. I forgot, I forgot to, yeah, usually when I do these, remembered it. They usually are like, and what's your name? So yeah. Monisha Saldanha (00:50) Cool. Okay, here we go. Here we go again from the top. Welcome to Colorado Tech People, the podcast where we talk with the founders and innovators building companies that shape how we live, work and connect. I'm your host, Monisha Saldana, an executive with 15 years of experience in product management. Today's episode explores how technology is transforming recycling. Joe Castagneri (00:53) Take two. Monisha Saldanha (01:13) I'm joined by Joe Castongherri, Director of Software of Colorado-based AMP, a company whose mission it is to modernize the world's recycling infrastructure by using AI-powered robotics to recover more valuable materials and reduce what ends up in landfills. AI for Good. We'll explore the hard decisions behind their innovations, the role of AI, how their technology is transforming everyday lives, and what it means to build and scale in Colorado. Let's dive in. Joe, delighted to have you here. Thank you so much for joining me today. Joe Castagneri (01:50) Thanks so much for inviting us on. Glad to be here. Monisha Saldanha (01:53) So let's start with just understanding AMP in a bit more detail. What was the problem that AMP was created to solve and what is the solution? Joe Castagneri (02:03) Well, put a little glibly, the problem is that it just costs more to sort some good stuff out of the trash than that stuff is worth. So we leave it in the trash. We bury it in landfill. And our answer to that is use a combination of computer vision, AI, and robotics to lower that cost of sortation so that we can get as much value out of our waste as possible. Monisha Saldanha (02:29) Wow, that's really simply put and what an exciting mission. AMP is tackling a deeply complex system, waste and recycling. What were the hardest early decisions about where to focus first? Joe Castagneri (02:43) Well, there's, If you ever tour around a recycling facility, which you'll hear me refer to as a MRF, a fun acronym, Municipal Recycling Facility, it's just a big old maze of conveyor belts and there are different pieces of heavy equipment that might take advantage of magnetism to sort steel cans or might take advantage of density to sort glass out from other things and you put the mixed recycling in or the mixed waste in and then on the other side you get commodities out. These facilities are typically fairly low tech by today's industrial standard. And so one of the difficult questions early on was not can we add technology into industrial waste spaces and find ways to increase efficiency. It was more where where should we what part of the problem. And really just like any venture where does the customer really see the value. Where do we actually help their business case if we make something easier. And it took a while for us to really figure out the incentive structure of this industry and where the value really was. Where we started was by building robots that take the places of the human sorters in these facilities. Because while you have all of these conveyor belts and pieces of heavy equipment towards the end of the facility, there's still this dull, dirty and dangerous job that in the past has been filled by people standing on the side of the conveyor belt and trying to pick 30 or so good pieces of plastic or aluminum or whatever off of the conveyor belt and put it into a bunker. And this is a job that people don't tend to stay in for a very long time. There's a high turnover problem. And so one of the first problems that we tried to tackle as an organization was, can we automate that role, build a robot that more reliably sees and then sorts the material off of the conveyor belt, can do it faster than people can and provide a positive return on investment on purchasing that robot for the facility operator. And for the first... several years of our existence we worked on that technical problem. So we built a computer vision system that looks at the conveyor belt material passes underneath the computer vision system. We use a neural network to say where and what the stuff is. And then that would tell at this time a pick and place robot. Hey be here at this time and use your suction cup gripper to grab this object and remove it from the belt. And early on the value proposition that we were really trying to provide was, can we get rid of this big turnover headache and stabilize labor costs and also stabilize the revenues that the facility operators receive from removing the good commodities out of the recycling stream. And that allowed us to work on our platform for several years. We built and installed several hundred robots across many dozen, a dozen, more than a dozen municipal recycling facilities in the US and some in Canada and even some in the EU and Japan. But ultimately what we found was that the that was a small part of their problem. It was just a small part of operating the facility to deal with this labor and commodity picking problem. Really most of their revenue ended up we found out came from receiving the material in the first place. Bulk of the revenue that a recycling facility receives is a tip fee, a dollars per ton amount that they get to just process material in the first place. So if you really want to help these facilities, technology that allows them to run more and run more tons of material reliably is a lot more value impactful to them. And once we figured that out, we started designing and building our own whole facilities that could use AI and robotics technology to build fundamentally more efficient recycling facilities and waste facilities. And that's what we do today. Monisha Saldanha (07:08) Wow, and how many of these facilities are there and where are they? Joe Castagneri (07:12) We built three. One is in North Carolina, operated by a partner of ours. And then we operate two. One is in Cleveland. And then the third is in Virginia. The Cleveland facility is a recycling facility. The Virginia facility is what we call a diversion facility. It's actually just sorting trash. It's not even resorting what you would call your single stream recycling bin. We're taking the black bin, just the unsorted waste, and we're pulling recyclables and other useful things out of it with our system in Virginia. Monisha Saldanha (07:50) Well that's really exciting and that's adding so much value to the environment as well as guess revenue for the facility. Why, What's happening in Louisville though? Your headquarters are here in Louisville. Joe Castagneri (08:05) Yes, so we have an office in the Colorado Tech Center. We've been here the whole time. So we've been in several buildings here in the Colorado Tech Center in Louisville over the years and have no plan to leave because we've kind of built ourselves the best laboratory that an engineer could hope for. So all of the robotic improvements that we're trying to make, we test out here. We also do all of our production manufacturing. You can imagine if you picture a big maze of conveyor belts. Well, if building a whole facility, there's a lot to build. And so we build and assemble and test all of that here before we send it out. And then exciting, we're actually building a recycling facility in Denver with Waste Connections right now. So we'll be, We've done all the design here in Colorado in this office, and then we'll procure and build all of the stuff for that facility. We are in the process of doing that. And then we'll install all of that into, it'll be up on Vasquez in Commerce City once it starts up later this year. Monisha Saldanha (09:15) Oh, that's exciting. That's coming up soon. And what are the future plans for expansion? Are you planning on more facilities across the U.S.? Joe Castagneri (09:26) Absolutely. Yeah. So we've got a pipeline of facility deals where we're talking to a whole bunch of different folks who you tend to see here are a combination of names that you'll recognize like Waste Management, Waste Connections, Waste Republic Services. They are the waste greats, the large publicly traded companies that manage waste hauling and land filling infrastructure. And they're always looking at how to modernize their infrastructure, their sorting facilities, how to make their landfills last longer. So we're in conversations with various folks at all of the waste majors. But then the other side are municipalities. so you, Like other utilities, your municipality often will have a board or a group who's in charge of awarding contracts to waste haulers for the purposes of managing the waste infrastructure. So that project I mentioned in Virginia, that was us working with the municipality directly that oversees about a million Americans' trash. And their problem was that their landfill was filling up too fast. And they need to divert material from landfills so that the landfill doesn't run out of space and then cause their their landfilling fees to really go up quickly. And so our solution there is building facilities that can divert half of what goes to landfill to instead use it for a better purpose. Recycle the plastics and in a carbon negative way remove all the organic material. Things like this that weren't possible before our sortation technology. entered the Monisha Saldanha (11:18) The carbon material, like the composting material, does it get sent to composting? Joe Castagneri (11:26) Good question. We're not composting it in that project. We take all of your food waste and yard waste. If you throw away a pizza crust in the Virginia Beach area, we'll sort it. We separate all the organic waste from the plastic recyclables, the metals, the paper, and then all the stuff we don't have a good use for. We take all the organics out, and then we are pyrolyzing it into a sort of charcoal. And what this does is it avoids what will otherwise happen, which is that stuff will get buried in a landfill and break down into methane. And landfills often have a methane recapture system. But a startling statistic, landfills are actually the third largest cause of human released methane in the world because of all of these pizza crusts that are breaking down buried underground. We are pyrolyzing it into basically a type of charcoal that you can use in asphalt and concrete. You can actually use it as a landfill cover to absorb some of that methane. But most importantly, all of the methane that would have been released by that organic material isn't getting released. And you have to go through a really rigorous calculation process to be able to claim, yes, this is done in a carbon negative way. And that's what we've done. And we actually recently announced a carbon deal with Google for, I think it was 200,000 tons of carbon dioxide removed over the next three years. Monisha Saldanha (13:07) What is the partnership with Google? Are you working with them on their waste? Joe Castagneri (13:13) So we've worked with Google in a variety of ways. Google's so big, I feel like sometimes you work with Google in seven ways and you've never talked to the same person twice. In this case, Google has a carbon credit group who is charged with finding carbon negative projects they can buy the credits for. And Google wants to offset their carbon impact. They want to be carbon neutral by buying carbon credits in the carbon market. And so they have a whole team that is identifying and acting on large carbon removal projects. And so that group is who we did this deal with, where they're going to buy the first carbon credits coming out of our project in Virginia. Monisha Saldanha (13:59) Wow, fantastic. And what happens to the actual carbon? Is it being put on top of the landfills in Virginia? Joe Castagneri (14:07) The current plant is using it as landfill cover. Yeah, so that helps filter both effluent coming out of the goo coming out of the trash stack, but it is a filter medium so it helps absorb the methane that is otherwise being released from the landfill as well. Monisha Saldanha (14:25) Does the carbon that you create take up less space in the landfill than the food the organic waste would take up? Joe Castagneri (14:33) Absolutely. Yep. And some high level stats because, you know, we're taking this wet goo and we're pyrolyzing it. You lose most of the mass, even just in water, right? All the water evaporates out of it. So it's a lot lighter and lower volume by the time you're using it as landfill cover in this case. Monisha Saldanha (14:42) Thank And what's the motivation of the Virginia municipality? Are they looking to save space in the landfill or reduce the methane that they release or both? Joe Castagneri (15:09) Their problem is a problem a lot of municipalities on the East Coast are facing. It's just that their landfills filling up too quick. And on the East Coast, there's not really room to make new landfills. And so if your landfill fills up, You now as a municipality have to freight your trash further away to a landfill that's not filled up. And at the least, you have to pay extra money for that additional freight. But generally what happens is that when the landfill supply is reduced in an area, the other landfills increase their fees. And so it's not uncommon if a landfill closes for you at your home with your trash bin out on the front for your trash bill to double or even triple. And so this, the municipal board that we did this, are doing this project with, they are looking out and are good stewards of the infrastructure and they want to make sure that in the long term, their member communities have access to affordable waste removal. And the key way to do that is to make their current landfill last longer. But it turns out that's a problem that's not unique to them at all. Landfills in general are filling up too fast. It's hard to make new landfills if we can divert material from landfill in a way that makes sense economically. It helps virtually every landfill on the East Coast, on both coasts. And of course, our goal eventually is to have it also make sense in places like Wyoming, where it isn't as expensive to landfill material due to land availability. Monisha Saldanha (16:53) And for the facility that you're opening up in Denver, what is the problem to solve there? Joe Castagneri (16:59) The problem there is that we have more people in Colorado producing more recycling. So there's just enough single stream recycling that waste connections is hauling around that it makes sense for them to have a facility that processes their material now where it didn't before. So they've just been scaling up in the Denver area. I think in part because there are more Denverites these days and also because recycling rates are taking up slightly in Denver, which is a good thing to see. So theirs is a good problem, but more straightforward. There's just more recycling demand as more people are moving here. Monisha Saldanha (17:40) And How did you decide to build a full stack solution combining AI, robotics, and infrastructure, rather than just focusing on one layer? Joe Castagneri (17:52) And one thing I'd say, I've been talking a lot about the facilities. And if we go to the beginning, we, the infrastructure part of all these conveyor belts and all the complexity of operating a facility, we had no idea that's what we were going to do. We were just building point, pick and place, pick and place robots that use AI to sort trash, which is already like fairly vertically integrated, but way smaller of a problem than the whole facility. So that made it a lot. Simpler where we started but then the other thing is that we were getting started, you know in the mid 2010s. and there weren't So there were really good off-the-shelf robots already you could get a solid Omron or ABB robot and then program it to do what you want it to and in fact, that's exactly what we did. But at the time there weren't good computer vision use neural network computer vision solutions that just existed in the marketplace. And so we knew and our founder and CTO, Matanya deserves the credit for this. He saw the opportunity that, hey, computer vision and neural networks and increasingly better and better Nvidia GPUs are going to make it possible to build computer vision perception models that are way better at seeing stuff than in the past. And that's going to allow us to automate things in industry that we couldn't automate in the past. And hey, the waste industry, the recycling industry is a little bit further behind technically than other industries are. So it's a good opportunity to try out that technology. So when you have that premise and then you say, OK, well to do this, we're going to to build a sensing suite. We're going to have to build a perception software stack that uses neural networks to see stuff. And we're also going to have to build a robotics stack to then go sort this stuff. What can I buy off the shelf to make this problem a little easier. And at the time the answer was, you can buy the actual robot off the shelf and the robot controller off the shelf. But at the time you had to build your fancy deep learning computer vision stack at home. So we spent a lot of time doing that and writing all the software that runs this neural network and then tells the robot where to pick and place. And then we did a huge amount of work actually getting it to work reliably. But we didn't have really another option at the time. We had to be kind of full stack with the exception of consumer off the shelf robots because nothing existed yet. Monisha Saldanha (20:36) Really interesting. And how has your approach on the software side changed with the advent of AI as we know it today? Are you able to find supplementary off-the-shelf solutions, or is your tech stack still wholly kind of custom-built software? Joe Castagneri (20:55) It really is majority custom built, but we've been very careful to set ourselves up so we can take advantage of how quickly everything is moving. So the deep learning tech space is a really big open source community where there are a lot of papers published that anybody can read. And so we've done everything that we can so that we can read those papers. And if it's like, that's better than what we got, we can rapidly prove it in our stack and then actually roll it out. Same thing with, you We use Nvidia GPUs because Nvidia is making way and way better GPUs every year. And so being able to quote unquote ride the tech curve has been really important where we have a pretty vertically integrated stack that we've built, but it's made up of these things where we can plug in new approaches that have been published in the kind of open source forums and new hardware that runs things faster so that we can take advantage of the benefits as they're coming. So a little bit mixed. We haven't though as an example outsourced our AI development. We still do all of it ourselves. We're just taking from what we're seeing the big research labs put out. Monisha Saldanha (22:20) And how do you find it running the company from Colorado? Like you mentioned that you're doing the R &D in Colorado. Is 100 % of your R &D happening in Colorado? Joe Castagneri (22:33) Yes, it's just we also have some folks that are full remote, so asterisk, but yes. Monisha Saldanha (22:40) And do you participate in the ecosystem in Colorado? Like do you find that there are, you know, either educational facilities that can support you or networking groups or other special interest groups that you can take part in? Joe Castagneri (22:58) Absolutely. I'm a proud CU alum, go Buffs. And our founder CTO, Matanya, also is. I started as an intern. And I found out about AMP because Matanya spoke at a CU event. Early on in the company's history, we were initially funded with an OED IT Advanced Industries grant. So there was some non-dilutive funding to basically just take R &D swings at Monisha Saldanha (23:03) you Joe Castagneri (23:26) Can you get a camera to see the trash at all? Is this worth even trying to build a prototype for? So it was very pre-seed, non-dilutive funding that sort of got us off the ground. We've had a lot of folks come from the local universities with different hiring projects that we've done with those universities. But then beyond that, there's a lot of networking that we take part in with groups like Colorado Clean Tech, which is a... a group that helps startups that are doing clean technology in Colorado find traction and find early angel stage funding. We have Colorado investors like Range as part of our backing. So definitely, I think there's like so many cool things about the tapestry of tech development in Colorado that we've been able to plug into. But it's also been exciting to see it grow because in the 10 years of AMP being around, we're seeing more tech investment and VC dollars flow into Colorado, which is exciting for me to see as a Colorado native, just to see more opportunities come up for this type of technological work. Monisha Saldanha (24:39) Great, and let's talk a little bit about hiring. So you mentioned that you do have remote employees as well who are outside of Colorado. How do you find hiring for tech talent within Colorado? Joe Castagneri (24:52) One thing that we find is just really strong mission alignment with Colorado applicants. A lot of folks in Colorado resonate with the idea of, it's a waste to bury that stuff in the ground that we went to all that trouble digging out. Can't we do something better with it? And so we find a lot of passionate and mission driven people in the Colorado pool. Not to say we don't elsewhere. Like it's a... People are allowed to want the environment to be better not in Colorado too, but I definitely find Coloradans seek out AMP in a higher number maybe than elsewhere Monisha Saldanha (25:34) Great. And going back to building the solution, was there a moment where AMP had to pivot the approach, either technically or commercially? And what drove that decision? Joe Castagneri (25:46) Definitely both. So I can give you two examples. Pivoting all the time. What I was talking about at the start where when you're in you're trying you're trying to introduce technology into a new industry and you think you know what their incentives are and how you can help like what they're going to see is value. And what we started with was this, we'll automate the human labor that is a difficult role that people aren't staying in for often the median of three weeks before people turn over in that role. Surely if we build a really good robot that does this, then we'll build a whole bunch of robots and everything will be done. And what we found, like I was mentioning before, is that the... 90 % of the revenue for these facilities is not coming from the commodities. It's coming from simply running. And so when we are a 1 % gain on the 10 % slice of their pie, it's just not really where the core value is for that industrial segment. And so that was one thing that ultimately led us to the commercial pivot of, we want to do greenfield facilities where we can show that this technology doesn't just enable efficiency and cost savings in the human sorting part, you can make a fundamentally better facility if you use AI and robotics as the Lego building blocks of the facility in the first place. And so that was a big commercial and technical pivot for sure. And I think the other side, well, yeah, I don't know. I think that I could give more examples, but that's definitely our... our biggest shift in focus has been from individual robots to the facilities overall. Monisha Saldanha (27:36) What's the decision that AMP made early on that felt risky at the time but proved critical to the trajectory of the company? Joe Castagneri (27:45) A big part of it is in that build versus buy calculus. You you're looking out at the market. Should we integrate this solution that exists or should we vertically integrate it by building it ourselves? And the example, The answer I would give is that we decided to build our own control system for these facilities. And that's it's a little bit niche, but the thing that actually turns all the conveyor belts on. It's how you turn the facility on and control the facility itself. Everybody who makes an industrial facility typically has a big industrial PLC, like a Siemens PLC that is doing this. And we sat there and said, well, we have these custom AI cameras that are looking at trash and saying what's there. And we have dozens of them. This is a fundamentally new kind of sensor that old style PLCs don't really know how to integrate with and can't get emergent value out of. We want to build our own. And it was a scary decision because everybody, you know, the normal advice is do only build the thing that's your thing. Like, buy what you can, buy whatever you can. But this would turn into our thing where we're able to do far more complex logic in our facilities than is possible on any other platform. And as an example, you can imagine just a security cameras looking at your facility running and it's we have a computer vision algorithm that's processing some of the video coming off that camera and it might be able to detect, hey, there's a jam starting to form on these conveyor belts. And that signal can go to our control system where we then maybe actually take a control action to not put material into that part of the facility and then flag it for an operator to go clear that jam before it really packs itself into something that is hard to clear and causes downtime which causes loss of revenue. That type of complex control logic just isn't possible unless you're fundamentally working on technology at the same level of your sensors. And so it's turned into a really emergent, powerful tech stack, but there was an 18-month period where we were like, please, trust us. We think this is going to be the right way. And it felt very uncertain at the time. Monisha Saldanha (30:21) And at a high level, how does AMP's AI actually see and identify different materials on a conveyor belt? Joe Castagneri (30:29) Honestly, lot like us, but dumber. So in a video, it's just a whole bunch of pictures, right? You you take a whole bunch of pictures per second, and each image is going through a piece of software I've been calling a neural network. And that neural network is taking in all those pixels and has been trained to identify the stuff in it. And what it's doing, basically, is drawing a box around every object and then classifying what's in that box. Monisha Saldanha (30:33) you Joe Castagneri (30:58) And those classifications are going to be things like milk jug, water bottle, aluminum can, and then one of our favorites, miscellaneous. And this type of neural network has a list of outputs that it can predict. So if I put my face under there, it's going to do its best and probably say miscellaneous. Like it doesn't know human. Human's not one of its labels. Monisha Saldanha (31:07) No. Joe Castagneri (31:26) So, you know, if every image is going in and it's drawing box around and classifying what's in that box, what comes out of the neural network now is the locations and types of materials and we can tell a robot to go pick that stuff. But how do we train it then? Because, you know, I've sort of buried the lead. What is the, how do you come up with this thing in the first place? And this is where it's actually a lot like us, but dumber. If I was going to train you, on how to identify all these different types of waste and recycling, I'd probably show you a whole bunch of images where I know the answers. And I'd say, what do you think this is? You'd say, I think that's a milk jug. And I'd say, close, kind of looks like one. And this one's actually a PET bottle. And then you're smart, so it might only take a couple dozen examples of all these different categories before you're good at classifying them yourself. With a neural network, we have to do it like 100 million times. So the training process is you just have these frames where you've labeled what everything actually is, show it to the model, have it guess, and based on how it's wrong, you tweak that model a little bit so that it's a little bit less wrong next time, and then you repeat. And that takes a long time, but at the end, you're left with something that is able to reliably classify and locate where the stuff is in the images. Monisha Saldanha (32:46) Yeah, and this kind of makes me think about the trade-off between accuracy and speed. As you mentioned earlier that with the larger facilities, really, you know, they can benefit a lot by processing more quickly. How do you think about the trade-off between accuracy and speed when deploying robotics in high-throughput facilities? Joe Castagneri (33:10) Speed ends up being the king because of what you're saying. Yeah, if you can run the number for these facilities tons per hour. So a pretty standard recycling facility will you'll feed it about 25 tons an hour of material. If you can feed it 30 then people will because that just means you can contract more tons to run and you get paid more. So but for us that trickles all the way down into you can only have the neural net take so much time to process each image. And you can't have too much latency in that decision or the material is going to be gone too soon. So these conveyor belts are all going three meters a second. So they're going pretty fast. It's like a full on run in human terms is the speed of these conveyors. And so we have our camera system looking at this stuff passing underneath them. And we only have a couple of feet of belt before we have to sort the stuff. And for this type of robot, we're using little puffs of air that the material will go over the edge of the belt and it'll puff a barrel blow it up to a higher belt if it is configured to or not blow a puff of air and it'll fall down and through and go to the next conveyor belt. So what we tend to find is that you have two types of constraint. You have a pretty hard constraint in how long you can allow your neural net to take or your latency of the decision. And you really have hundreds of milliseconds is your envelope there. And that limits how heavy of a neural network you can run. Also sort of requires that what you're running is local. You can't go to the cloud at those latencies. You just don't have enough time. And then the second constraint is that, it's got to be good enough at what it does that it makes something you can sell. You have to have high enough purity material sorted that it actually is any good. And those two constraints can battle each other a little bit because how do you make the quality better? Maybe a heavier, smarter neural net. That'll be a little bit slower. So then we do work to make it run faster if we can. And a big moment in the 2010s is that neural networks reached a point where they were capable of both of those at the same time because neural networks have actually been around for decades, but we did not have the compute to run them very fast. GPUs weren't mature and didn't exist in the 90s, which was a previous wave of neural network stuff. But, these things existed. It's just like, sure, maybe it can come up with an answer. It's just going to be in in an hour. And it needs to be in 100 milliseconds. And so I think the other side of this question is the reason AMP exists now is because of the AI wave reopening in 2011 onwards with convolutional neural networks and GPUs really making this heavy of an algorithm runnable faster enough to begin with. Because you need this heavy of an algorithm for it to be accurate enough. Monisha Saldanha (36:00) you Yeah, and the recycling environment is messy and unpredictable compared to, say, a production facility in manufacturing. What makes building AI for this use case fundamentally different for more controlled environments? Joe Castagneri (36:43) The, I call it the fat tails problem. If you think about a bell curve, you know, you have your normal stuff in the middle and then you have the tails. And for something like building computer vision perception in a manufacturing environment, like you're saying, your tails are really narrow because it's all well defined. You're only ever going to see if you're assembling a car, the parts that go into that car. But in a trash facility, you're eventually going to see somebody's stuffed animal that's been beheaded, you're going to see a full bicycle, you will eventually see an anvil, and you have these really rare items that come up very commonly. It's like each individual item is rare, but it's not rare to see a weird thing. So that makes it a hard classification problem. It's also a subtle classification problem. The difference between a piece of white paper and a piece of shredded white plastic can be really difficult to tell, but they're completely different materials. Contrast that to like self-driving cars. The difference between a bicycle and a bicycle on the back of a car is subtle, but it's less subtle than two white, scrappy scraps that are both dirty. So it ends up being a really difficult classification problem where we need a lot of data and really, really high quality training data where we've gone through and meticulously said what the boundary is between this and that. And so we end up having to spend a lot more time than other perception teams would on the dataset quality and defining what is what. We have an incredibly talented data team whose whole job is to build up that data set to be as accurate as possible so that our model can be accurate enough. But the problems don't end in AI. Because all that stuff is there, it's physically there, it also is really hard to build robots that can stand up to it. It's a really dirty environment, lots of dust, lots of grime, lots of... impacts from the anvils that break stuff. And so it didn't actually take us all that long at the very beginning to have our first demo of like, here's a robot that can sort trash. It can see it and it can sort it. But we put that into our first recycling facility, which was in Denver. It's what's the current GFL facility in north of Denver. And it took about seven minutes before it broke. because it's just a hard environment to mechanically exist in. And so a lot of our work in the first five years of the company was not just getting it to work, but getting it to be reliable enough to work in this really brutal environment. And it taught me a lesson of like the thing you think is gonna be the hard thing isn't the hard thing. Something like the hose wearing out quicker than it needs to wear out is gonna haunt you for three years and it did. Stuff like that. What you think is gonna be easy is gonna be hard. Monisha Saldanha (39:57) That's a good lesson. Recycling can feel abstract to many people. How does AMP's technology tangibly impact everyday lives? Joe Castagneri (40:09) One of the keys is that we want to make it easier on the consumer. As a person who's worked in this at AMP for almost 10 years, my friends are always, what do I do with this? Am I allowed to put it in that bin? It feels like the rules are always changing and we always get it wrong. And we want to make it easier. So in that Virginia facility, we're just sorting the trash. A couple of cool things come out of that. First cool thing that comes out of that is that if you live there, you don't need to have a separate recycling bin so you don't have to think as much. Right. You don't have to think like does this tub belong here or there. It's way easier to not make a mistake when you only have one bin. But the key is that it's actually increasing their recycling rate tremendously. So the current recycling rate is only 6 to 8 percent in those municipalities and our facilities are going to bring it up to more like 20. And that's because it's just we're taking the recyclables out of the trash already. We don't need to have access to the recyclables like you do in a single stream program. That being said, single stream programs matter a lot too. So we want to make single stream facilities more effective at bringing, pulling out more of the valuable stuff in there. So again, you don't have to think as much. If it has the symbol on it, we'd like to be able to remove it. And so it's less complicated to use as a consumer. Our end goal is to make it so as a consumer, you don't have to think at all about it. We're getting the value out of the trash because the value is there and worth taking. And that's what we want out of these diversion projects. Monisha Saldanha (41:54) So could you imagine a future where there is no more single stream recycling and everybody all across the United States just has the one bin that everything goes into? Joe Castagneri (42:06) I see that world and I think it's quite a ways away for some areas versus others like the areas that there currently isn't recycling infrastructure. That's where we want to start. Right. Increase the recycling rate by putting a diversion facility there because they can afford a diversion facility but a full recycling facility might not make economic sense. That brings up the recycling rate in the immediate term without having direct competition with the recycling facilities. Let's go to the markets where recycling facilities are denied first. But yeah, in the long term, if the technology is good enough, then yeah, we'd love to reduce costs for municipalities by reducing how many hauling systems they need to have. have Denver has three bin system, compost, recycling and black bin waste. Well, you need three trucks and three routes and logistics for all three of those. That's expensive. And so, yeah, our goal would be long term to be able to reduce expenses to municipalities on the hauling side and increase recycling rates on the actual sorting side by simplifying the whole thing to one big mongo facility. Monisha Saldanha (43:25) Exciting vision. Final question for you. What is one book every builder should read and why? Joe Castagneri (43:33) Well, can I cheat? I have two answers. One's fun. The fun answer is there's this book called Takedown, The Fall of the Last Mafia Empire that's about how the mob used to control trash in Manhattan, way more recent than you might think. It was through the 90s. And it's really interesting. It's kind of to see how this industry Monisha Saldanha (43:36) Yeah, you can have two bucks. Mm-hmm. Joe Castagneri (44:02) the history of this industry and its ties to the mob, especially on the East Coast, which have been largely broken now, which is great, but it's just kind of an interesting book if you're interested in waste. It's not every day I'm on a podcast to talk about waste and books, so I figured I'd had to pitch it. But if nobody you've talked to has taken the Innovator's Dilemma yet, I think that would be my recommendation for what builders and founders should read. The Innovator's Dilemma defines what we mean by disruptive technology versus iterative technology. And it expounds on this thesis that the reason that incumbent firms don't invest in disruptive technology is not because they're dumb, it's because their revenue incentives make it really difficult for them to. And as a consequence, the technology development ends up happening by way smaller players that can put all of their focus on the technology that they're building because all of their revenue is coming from it. And that makes them really dangerous to those incumbent firms that are going to have more trouble establishing the focus and working on the new technological innovation. And this book goes through, you know. 50 examples from hard drive development and computers all the way to business models like how the way that we did retail clothing sales really changed business models in a fundamental way that caused a lot of incumbent companies to fail. And I think it's important for builders and founders because it helps you understand where incumbent firms have weak points that are therefore opportunities for new ventures. If there's something that the incumbents aren't going to be good at solving, well, that's something where maybe you have a chance to solve it for them in a way that could bring value to the market. So innovator's Dilemma. Good book that made me think a lot. Monisha Saldanha (46:11) Fantastic. Thank you for that. I have read The Innovator's Dilemma, but I have not read the first book that you mentioned, so I will take down. Joe Castagneri (46:19) Take down, yeah, yeah, if you, it's kind of fun. And it's a true story. It feels like historical fiction, but it's real. Monisha Saldanha (46:24) Yeah. Fantastic. Well, Joe this has been a great conversation. Thank you so much for joining me today. Joe Castagneri (46:37) Of course, and thanks again. We're happy to be on. It's exciting to see this podcast as an example. There's enough tech happening in Colorado that there are conversations to be had about it. So I'm glad to be on. Monisha Saldanha (46:51) And I'd like to thank our listeners for listening to this episode of Colorado Tech People. If you enjoyed our conversation about using technology to improve recycling outcomes, consider sharing this episode with someone who loves entrepreneurship. Be sure to subscribe so you don't miss future conversations with founders and leaders shaping Colorado's tech ecosystem. Until next time, keep building, keep connecting, and keep creating experiences that bring people together.