Interview Transcript

Disclaimer: This interview is for informational purposes only and should not be relied upon as a basis for investment decisions. In Practise is an independent publisher and all opinions expressed by guests are solely their own opinions and do not reflect the opinion of In Practise

We've been investors in Amazon for nearly 10 years, but our understanding of automation and robotics is quite limited. We've focused on other business areas. The purpose of this research project is to understand the state of intelligent automated systems and robotics within leading companies. Let's dive in. The first question I have, which I think is a good starting point, is about how automation and robotics have transformed Amazon's e-commerce fulfillment operations during your time there.

It's important to understand that warehouse automation varies significantly depending on where you are in the supply chain. For instance, in the mass goods area, most items entering the country arrive in containers. These containers are typically sent to a location where they are unloaded and placed on pallets, which is the most common practice. The pallets are then shipped to warehouses.

My involvement with warehouse automation began when I was hired by CNS Wholesale Grocers, the largest independent wholesale grocer in the US. Working directly for them provided me with a real education in warehouse operations. Previously, my experience was primarily in manufacturing, not warehousing.

If we look back about 30 years, the first type of automation introduced was in pallet warehouses. At that time, the majority of warehouses were not automated. The typical setup involved a voice picking system. Warehouses were organized to be efficient for manual pickers, who would drive double pallet trucks up and down the aisles. The voice pick system would direct them to specific locations.

When a picker reached an aisle, they would verify their location by reading a license plate and confirming it in the system to ensure accuracy. The voice system would instruct them on the quantity to pick, usually one or two items, which they would place on the pallets before moving to the next picking location.

The most important person in the warehouse is what they call the "slotter." The slotter decides where to place items in the warehouse. The main objective, especially when shipping to stores like grocery stores, is crucial because these stores often have almost no backroom inventory. Most of what's coming off the truck goes directly to the floor. Therefore, you need to have all the same items, the same SKUs, on one pallet.

On a lighter note, my first job in high school was at a grocery store. I understand that if a pallet had pasta sauce, baby food, and detergent, it would require manual labor to distribute the items correctly. In-store labor is essential. In a manual warehouse, you're trying to arrange your warehouse aisles to resemble what the store receives.

Our biggest customer at the time was Ahold, which includes Stop & Shop and Giant Eagle, one of the larger chains in the US. Each store would have a different planogram because stores vary in size and community mix. You're trying to layout the warehouse to efficiently serve the stores, determining where to place items for efficient picking.

You always want to maximize the cube height of the warehouse. Typically, warehouses are 40 feet tall. When I look online at other systems, one of the first things I notice is how poorly they utilize the full cube. Items are often placed on the floor or at a readily accessible level, leaving unused space above, which means the warehouse cube isn't fully utilized.
Fundamentally, a warehouse is for storing items. As the number of SKUs increases, particularly in e-commerce, space becomes limited. Initially, warehouse automation systems were crane-type systems. A pallet would be placed onto a crane at the end of an aisle, which would then move down the aisle and lift it to a level. These systems often had a load handling device on the crane that automatically placed items onto the shelf. This technology is 20 to 30 years old and originated largely from Europe.

In the US, the leading supplier was HK Systems, a supplier of CNS, which was eventually acquired by Dematic. This is essentially where warehouse automation began. Moving forward, the next step was automating the picking system I described. For example, when replenishing a big box retail or grocery store, you're mainly dealing with cartons or cases, such as cardboard boxes, shrink wrap trays, or sometimes bags like dog food. These items would be manually picked by the store picker, requiring them to build a Tetris-like puzzle on the pallet.

The first successful examples in this area were in beverages, like beer and soda, where even with 500 or 1,000 SKUs, many boxes were the same size, making pallet or case handling relatively easy. However, in a grocery store, there's a vast variety of packages.
Solving the puzzle software was a challenge for the industry. The first company to succeed was Witron, based in Parkstein, Germany. They had a PhD scientist named Dr. Lee, a German-Chinese gentleman who developed the first capable puzzle software for building decent pallets.

The puzzle problem is a classic mathematical problem. In mathematics, there are levels of problems that are solved and unsolved. An NP-complete problem is an unsolved problem with approximate solutions, like the traveling salesman problem. A step up is NP-hard, which is more difficult.

The knapsack problem is an NP-complete problem taught in computer science algorithm courses. It involves determining how to fit the most items into a container of a certain size. Our challenge is even harder because it also requires stability, ensuring items don't fall over. Most algorithms at the time could start building a decent base but would eventually run out of good options, leading to the creation of chimneys, where cases of the same shape are stacked on top of each other or become smaller.

You end up with something that's highly unstable. It's unstable in the warehouse when you're building it, and even worse, it's unstable on the store floor. Imagine trying to take apart an eight-foot-high pallet—it's a mess. That was around 2008, which is when I joined Symbotic.

Rick Cohen, the owner of CNS and Symbotic, was someone I reported directly to. We got quotes from Witron. I know about Witron's technology because I visited their facilities in Germany and their sites in the US to understand how their system worked. They had functional software that worked effectively.

The second thing they did was build something called the COM machine, a case order module. They stored cases, which come in various shapes and sizes, on trays similar to those you'd find in a cafeteria. These trays were easier for automation to handle, allowing them to be placed into shelves.

Think of a tray as a tote without sidewalls. They built enormous warehouses, starting with a pallet warehouse for bulk storage. As needed, they would depalletize and transfer the cases onto trays, which were then stored. A crane system would pick the trays in the required order.

The challenge is that to build the pallet robotically, you must bring the cases out in perfect sequence. If you're building the pallet and case number 20 isn't there, you can't place items on top of it because it will fall over.

They built a case order module, which was essentially a giant German-engineered sorting machine. It required a lot of space and was costly. At that time, you also had to construct a dedicated building because retrofitting an existing one wasn't an option. The automation alone cost between $80 to $100 million, not to mention the time and expense of building the warehouse.

They had a customer, Kroger, in California, where they built one. Consider the time it takes to get permits in California and the size of the project. This was around 2008 when CNS decided to invest in case picking. I was hired to conduct R&D for CNS. My main responsibility was to define the milestones because Rick, who was being asked to invest $17 million by the inventor, had never invested in a startup before.

I explained to Rick that instead of giving all the money upfront, he should break it down into milestones. There were six milestones, and I defined them. I reported back to Rick that the inventor, John Lurch, was like Doc Brown from Back to the Future. He had big ideas but never made anything work. My job was to coach him, and I told John that my role was to help him succeed, but I had to approve the milestones.

John failed on milestone five, and we discovered he had falsified some results on milestone four. We had to decide whether to shut them down or acquire them. We chose to acquire them, and that's how Symbotic was born. I then moved in as the CTO. We developed even better puzzle software ourselves. We explored most of the software available at the time.

Dr. Lee had moved from Dematic to Schaefer, and they both used the same puzzle software. CNS actually bought Schaefer's first palletizing system. We considered Dematic and other players at the time, but they all faced similar challenges. We then approached a company called Tops, which was an independent software provider specializing in truck and container loading.

When we reached the production stage at our first site in Newburgh, New York, we had a live, fully automated warehouse. However, we could only build pallets up to about 5 feet tall before their algorithm failed. We learned a lot from this experience.

One of your questions was about our most significant accomplishments. We developed the best palletizing software, which might still be the best, though others are catching up. Initially, we thought the key was to learn from the best pallet handlers and incorporate that into the software. I remember videotaping workers with double pallet jacks to observe their methods. While informative, this approach was ultimately a dead end for us and others.

Human problem-solving in complex three-dimensional spaces is difficult to translate into rules or AI. We got really good at palletizing software by using mobile robots to shuttle up and down aisles and then transition to what we called the dance floor in AMR mode. This allowed them to organize items into the correct order. We eliminated the need for what Witron had as a comm machine by handling everything in software, which gave us a significant advantage.

If you consider the Witron example or any scenario where items are stored on trays and totes, there's an inherent inefficiency in storage density. This is due to the space around the box; for instance, if you place a small box on a large tray, there's air all around it. Additionally, if you have shelves that are all the same height, you'll end up with boxes that are too short, leaving air space.

To address this, we developed a system where we don't use trays. Instead, we go directly to the boxes, allowing for a narrow tray next to a wide tray on the same shelf, minimizing the space in between. We also implemented shelves of varying heights, accommodating both tall and short boxes. We wrote computer software to optimize the warehouse design based on the mix of SKUs in inventory.

As a startup, we managed to break into Target. They were expanding their warehouse in Woodland, California, near Sacramento. They had previously purchased four systems from Swisslog, known as the Atlas system, which had gone through growth pains but eventually worked well. However, the state and other bidders proposed spending $60 to $80 million to extend their existing warehouse, which would be time-consuming and costly. We were the only ones who could fit the entire system into their existing building by removing their old, inefficient pallet rack.

When Target asked other vendors to re-quote, none could match our proposal. This is how we, as a startup, with Rick Collins' financial backing, succeeded. Our system offered the best storage density, perfect sequencing, and advanced software capabilities. It took us four years to build the first fully automated facility in Newburgh, which was dock-to-dock automated. We received inbound full pallets, integrated them into the system, and produced shrink-wrapped, aisle-specific pallets for outbound shipping.

We learned a lot from building our first system, including identifying mistakes. We received orders from Coca-Cola and Cisco Food Service, who saw us as a threat. We expected them to lower their prices, so we aimed to reduce costs. Rick challenged us to improve storage density by 40%, increase cases per hour by 40%, and reduce costs by 40%. We called this System X and spent a year doing a lot of simulations on optimization, taking a cross-functional approach.

Unlike many companies, we put everything on the table, changing hardware, software, controls, and building structure. We went through numerous iterations, and I led the effort, resulting in 35 patents, mostly related to System X.

When you ask about the most significant thing, it was System X. There wasn't just one thing, but those 34 patents cover all that. That's the system they're selling to Walmart today. You can see it online. It's basically the same system that was there when I left in 2015.

I had lunch with some former teammates about a year ago. I had hired them away to go to Amazon, and after Amazon's big layoffs, they returned to Symbotic. On the Symbotic website, everything is labeled as AI. I asked them about it because I didn't recall any AI back then. They just laughed, saying it's mostly a marketing thing.

I spoke at the Boston Robot Summit last April about the ARM Institute. A guy from Symbotic, hired after I left, talked about their technology and AI. I was impressed. He focused on machine learning and pattern recognition, which are the most feasible areas.

One of the biggest issues in a warehouse is the variability in the quality of packages, cartons, and shrink wrap. If a robot gets stuck, it's usually because it can't handle the package.

What do you mean by "can't handle the package"?

The package might be defective or have free flaps.

It's torn or not right. The robot is programmed to pick up a box in a certain way, but if the box is broken, it won't be able to.

We began to encounter issues early on, so we developed basic sensors to detect significant problems and instruct the robot not to proceed. In warehouses, you store SKUs, and you don't want them all in the same place. You want to spread them out across the warehouse to have multiple picking locations. This is not only to prevent damage but also for efficiency. For example, if someone orders four items, you don't want them all in the same aisle because only one robot can go down an aisle at a time. You want to pick four items simultaneously from different aisles and then bring them together to place on a pallet.

What was being done, although still experimental, involved using images to detect more defects in cases. These defects could determine whether to pick an item or not. During off-hours, when picking isn't happening, a rescue bot or a person might handle it differently.

At the same Boston conference, I met someone from Procter & Gamble who oversees all vision applications globally. They had 60 or 90 installations worldwide using machine learning to detect defects. For example, consider a plastic twist bottle with a label on the back. If there's a wrinkle crossing the barcode, it can't be read. You want to detect that before it reaches the store. They have the advantage of having more data than needed for training.

I've heard that a major problem many applications face is the difficulty in obtaining enough data for training. Let me give you a counterexample. The ARM Institute had a really cool project, a joint effort between Boeing, Carnegie Mellon, and Siemens. They were working on gluing fiberglass carbon fiber panels together. Along the edge, they needed to place glue drops and wanted to visually inspect these drops. However, conducting the experiment and gathering sufficient data was very expensive.

In this project, I believe Nvidia was also involved. They aimed to create synthetic images based on the few glue drops they had. They artificially generated drops with defects, like a little drip sticking out the side. They collected a number of defects and used them to generate more data instances artificially. Visually, it was hard to distinguish between real and fake drops, especially with Nvidia's involvement, as the images were super realistic. This example highlights how challenging it is to validate data without enough of it, and by the time you reach production, it might be too late.

Now, regarding systems handling shuttles, totes, or trays, I mentioned the storage density problem. There's a bigger issue because most ecommerce, except for Amazon, stores items in totes, like AutoStore and others. The problem arises when you start with a full tote, and as you pick items, it becomes only one-third or one-quarter full. This affects your storage density. You can calculate theoretical storage density if everything is full, but with partially filled pallets, the effective storage density is much lower. I noticed this when I walked through a Target warehouse and saw partially picked pallets, resulting in poor effective storage density.
The solution is to consolidate your inventory. For example, if you have the same SKU in four different locations and they're all partially picked, you should bring them together into a consolidated tote. In grocery, it's more challenging due to date codes, so you must be careful. Consolidation requires picking items and bringing them to a place where either a robot or a person can consolidate them. This process consumes a lot of automation capacity, which doesn't add value to swift fulfillment.

Before I move on to Amazon, do you have any questions about what I've covered so far?

No, I'm sure we'll delve into it after discussing Amazon, but this has all been fascinating so far.

Amazon's significant move was acquiring Kiva roughly 12 years ago. I remember Symbotic was just 15 minutes away, and we were surprised that someone would pay millions for anything in this space. What people underestimated was that, unlike most startups, Kiva had a single, massive customer who would bankroll them, keeping them focused on one thing.
It wasn't as focused as you might expect, but it was more focused than many warehouse automation startups, which often chase multiple objectives, diluting their efforts. Besides the financial backing and market access, Amazon brought an enormous software capability.

When I joined in 2017, there were a couple of thousand highly talented software engineers in a group called Amazon Fulfillment Technology (AFT). This group developed the software that ran in the cloud, directing all the automation. People underestimated the power this would bring.

At that time, we were open about public tours and videos on the Internet. You could walk in and see all these yellow totes moving around, but without understanding the software behind it. I was on the same management team at Amazon Robotics that handled the Kiva technology. There was also the Advanced Technologies group, which dealt with industrial robot-based tasks and packaging.

AFT was involved, and there was a group called Worldwide Engineering that managed the eight miles of conveyors and sorters, handling material logistics. Initially, I was heavily involved with procurement contracting with EMHE, which we can discuss later.

I took over the group responsible for all installations, commissioning, and deployments at the sites. Everything was reporting to one person, which gave me a comprehensive view of the operations.

Each year, like most companies, we had to demonstrate how much money we saved the company. This needed to be audited by finance; you couldn't just estimate savings. You had to show where it appeared in the P&L, which I believe is essential.

One year, AFT saved the company a billion dollars. It wasn't just one thing; it was a combination of factors. The automation was powerful because it enhanced human productivity.
One of your questions is about the extent of automation. Depending on how you count, an item might be touched 12 or more times from when it enters a fulfillment center until it is on a truck. Only a few of those processes are automated, which is why you still need 2,500 people in the building, even in an automated facility.

I'll explain what's happening now and what they're working on to address this. The KIVA system was effective in reducing walking. There are picking and stowing stations surrounding the area. On each of the four floors, there's a cage. We refer to areas as inside the cage and outside the cage. Inside the cage, KIVA robots move around with storage pods, distributing inventory. The goal is to avoid placing the same SKU in the same pod, providing multiple choices when fulfilling orders.

When the inventory comes in, 80% to 85% of it arrives from what they call receive centers. When the items come from the manufacturer, either domestic or international, they arrive on pallets. These pallets are emptied into yellow totes and then fed into sorting systems from a company called OPEX. This system is specially designed, based on postal sorting, as OPEX originated from postal sorting and sorting envelopes.

The system allows you to put items from a tote, for example, all toothpaste, into the system, which then distributes them to around 16 different tote destinations automatically. This process is mostly touchless. Each destination is another fulfillment center. By the time items reach the fulfillment center, they are already sorted down to the item level. However, in those totes, you have a mix of items, creating a "fruit salad" from the single-item totes entering the OPEX machine.

These mixed totes are sent to the fulfillment centers. The other piece of automation involves handling the large number of totes coming out. These totes are easily stackable and palletizable. We brought in hundreds of Fanuc robots to build pallets at the cross-dock centers. Now, they enter Amazon, and conveyor systems transport them to different levels. At the stowing stations, people take items out of the totes. A pick-to-light system guides them, as the system knows what's in the tote due to the barcode label.

The system instructs the operator on which item to pick from the tote, with a computer screen displaying the details. The operator then barcodes it to verify it's the correct item. Next, he moves to the yellow pod, where the computer suggests options for maximizing density. The operator selects where to place the item and barcodes the label beneath that storage area.

One of the significant productivity advancements during my time there was the use of machine learning to reduce the number of items requiring barcoding. Although the job remains the same, machine learning was employed to ensure, with high confidence, that the right item is placed in the correct location. This is the stowing part, which is rarely shown online due to its messy nature, involving packaging and dunnage.

On the picking side, the process is reversed. At the picking station, the computer instructs the operator on what to pick and from which storage location. They intentionally differentiate items in each picking location to prevent errors. For example, back when CDs were common, you wouldn't fill a location with multiple CDs, even if they were different, to avoid picking mistakes. Instead, you'd have just one CD or one toothpaste to minimize errors during picking.

The tote isn't picking an order; it doesn't know what that is. If you average three items per order, they come from different places in the warehouse, picked by different people, and placed into totes. These totes are mixed because they are destined for the order consolidation area.

Once in the order consolidation area, an individual picks items from the "fruit salad," barcodes them, and places them into individual trays. These trays move to the other end, where a person takes the item out. The system knows what the item is and the order it belongs to.

Behind this person is a "put wall," a set of giant cubbyholes. The cubbyhole lights up, indicating where to place the item. When the order is complete, a light on the other side lights up. For multi-line orders, a person on the other side puts them into a box or bag. Single-line orders go directly to an automatic packaging or bagging system.

I've described the process flow. What people often overlook is the role of the software, which is something you can observe by walking around. The key is the limited number of large cubbyholes in the put wall. You don't want a partially filled order occupying a space for too long, as it clogs the system. You need everything to come out more or less at the same time.

It's not precise sequencing like we had to do with the puzzle software for the pallet, but it's close. You need to manage all the components coming down.

Another interesting aspect is, as I mentioned, there's a person taking items out of the fruit salad and placing them on a conveyor for another person to take off at the other end and put into the foot wall. They became very sophisticated in understanding the number of pieces per minute or per hour that each of these two individuals can handle. You want them to be balanced. If one person is consistently faster, the slower person ends up blocking or starving the faster one. This is a function of time.

They became very adept at analyzing individual performance as a function of the time of day. The same applies to those doing the picking and stowing. The productivity gains came from optimizing the use of personnel, considering the manual labor involved, and synchronizing everything efficiently.

Another important aspect is inside the cage where all the Kiva robots operate. If you notice, there are almost no aisles in that storage area because aisles reduce density. They developed software that could efficiently move items around to access the ones needed, even if they were in the middle of a parking lot. This was a challenging task.

Additionally, you want the pods to arrive at the picking and stowing stations so that the person is always busy, never waiting for a pod to arrive. There should always be one ready to go, and you don't want a pod blocking the way out.

When I mentioned there's a billion dollars, you might wonder how that's possible. If you consider those 1,500 to 2,500 people at peak doing that work, and multiply by the number of warehouses, you can make a change that improves efficiency by 10, 20, or 30 pips per hour. That's how those numbers come up.

Around 2019, before Covid, Amazon decided to make the next leap to one-day or two-day delivery. At that time, their network relied heavily on UPS, FedEx, and the Postal Service for downstream services, which they called sort centers. Over time, FedEx and UPS decided not to keep investing in buildings solely for Amazon, so Amazon started building their own facilities. However, achieving one-day or two-day delivery nationwide was challenging if multiple stops were involved.

If you're picking an order, sending it to a sort center by truck, sorting it, and then putting it on delivery trucks, it's hard to make it all work in time. So, they needed to eliminate steps and move buildings closer to customers, which is difficult in urban centers. At that time, fulfillment centers used cross-belt sorters, which are essentially racetracks 30 feet in the air. Packages were placed on carts with conveyor belts, sorted, and dumped into bags or large cardboard boxes called Gaylords.

The limitation was fitting only 60 to 80 sort points in a building. That's when they transitioned to Pegasus, the blue robot with a conveyor on top. Instead of a racetrack, 200 to 300 Pegasus robots could run around doing the same job, dropping packages down chutes to bags or Gaylords on the floor below. This technology was deployed around 2020-2021 during Covid. My team built and installed these systems, experiencing growth pains before they were ready.

This was primarily a transportation lead time and cost play, not a labor play. Amazon's key productivity metric is VCPU, or variable cost per unit, which measures the cost to deliver an ordered item to the customer. Transportation cost is a major factor in this metric, influencing many decisions. If the cube density of shipping methods like totes, delivery vans, or trucks is poor, it's a disadvantage.

The reason this is interesting is that Amazon began using what they call go-karts. These are wireframe structures with multiple shelves and wheels, which were packed manually and then wheeled onto the truck. The advantage of this method is the speed at which you can load and unload, compared to using cartons or bags. Fluid loading involves a conveyor running into the truck, and manually building that up takes a long time. Most warehouses are limited by the number of dock doors they have, so this ties up a truck and a driver, creating inefficiencies.

Go-karts have the advantage of speed, but you wouldn't want to ship them across the country due to poor storage density. However, in urban areas, they could be beneficial because the impact of having air in your truck isn't as significant. They started with this, but the application was somewhat limited, which led to the development of what they call Proteus. These are lime green, true AMRs. I was in the same building when these were being developed, and my team worked closely with that team.

If you look online, Nvidia has some really cool videos of this project, showing the AMRs in action and the 3D imagery of the real warehouses they operate in. My group contributed to this by providing the ability to detect obstacles not included in the engineering designs, like sprinklers. Sprinklers are usually installed at the end of the building by a contractor to meet fire codes, but their exact locations aren't always known.

We could detect that with later IR scanning. The idea here is that with this system, you have the sorting robots. There's a chute, and then a Proteus robot comes over to pick up the Gaylord or whatever the destination vehicle is. It takes it to another robot called Cardinal, which is much bigger. The Cardinal robot takes the packages and places them onto the shelves on the go-kart, and then the Proteus can then take the go-kart to the dock door, ready to be loaded onto the truck.

This was a big deal, partly for labor reasons, but more so for safety. The outbound dock is the most congested area and has the highest incidence of accidents. There's a lot of publicity about Amazon and safety, so this was a significant issue to address. It was a challenging problem to solve due to the congestion and unpredictability, with obstacles in different places at different times, human beings walking around, and forklifts operating.

They have just launched this system, and I believe it's now at two sites. Nashville was one, where we did a lot of the pilot work. They just announced a more automated site in Shreveport, which I assume includes this system. Another major change, this is all public, by the way, is that until recently, they were using the yellow storage pods from the Kiva days. Inside a Gen 1 of those Gen 11 buildings from the 2021-2022 timeframe, there were 100,000 of those in one building. It was uneconomical to build them in a factory and ship them. So, what we did, my team, was bring in all that as raw material, like cold-rolled steel bars.

We assembled it all into a frame structure and had to build a portable assembly line to do that. All the fabric came in on a pallet, all stacked up. In about a 14-week period, we built 100,000 of those on-site across four different floors.

Did you subcontract that out to fabricators?

No. One of the questions was about the difference in our ability to scale. We managed to do something that no one else did, and I was amazed. I couldn't believe it when I got there. We self-integrated when we purchased equipment. We were the largest customer for Intelligrated, Thematic, Vanderlande, and others. I negotiated those contracts and worked with those companies when they brought in their teams to install the sorters and conveyors. We handled all the system integration, physically connecting everything together.

Additionally, since all the software is in the cloud, we developed our own warehouse automation and warehouse control system, which we call the AWCS, Amazon Warehouse Control System. It served as a standardized interface layer between the cloud software and the hardware. We developed our own proprietary interface standard, which we required all vendors to comply with, instead of using their usual methods for other customers. Initially, there was pushback because they wanted to use their systems, but we insisted on uniformity across our warehouses.

Before Covid, we would send guys out with their laptops to do all the configuring and testing, including cybersecurity. During Covid, we couldn't send our teams to sites, so we developed more automation and scripts to configure and test systems remotely, with only one person on site. Eventually, we created a self-service app for this purpose. If you asked those vendors, they would be unhappy if you took that app away because it significantly reduced their time on site, which was crucial during Covid.

We worked with third-party assemblers, or 3PAs, who handled the physical labor, such as stretching fabric over frames and riveting them together. We provided standard work instructions on tablets, complete with training videos and step-by-step checklists, allowing us to track progress. We developed our own logistics systems to manage the delivery of materials, receiving up to 11 53-foot trailers of material daily. That's how much material was coming into the building.

We ran our operations like a factory, with a lot of automation. The use of 3D LiDAR scanning reduced change orders and saved the company significant time and money. During the challenging Covid year of 2021, we built around 80 greenfield buildings of various sizes and completed approximately 250 upgrade retrofit projects in one year. Our metric was a 98% or better on-time delivery rate.

That's amazing.

We had a 98% or better rate of no high-severity tickets in the first 30 days after launch. The biggest problem was with our suppliers. They worked hard, but their supply chains were deficient. They had no systems. I visited one company on site because they were struggling. I asked, "Where are your work instructions?" They replied, "Maybe there's a binder on the trailer." They had work tables with paper drawings but no systems.

We offered to help them build these systems. We pointed out that their products and instructions would differ with the tools we use. A building delayed by a day in launch costs a million dollars. During Covid, conveyors we needed were four to six months late. We constantly juggled to manage that. One differentiator between Amazon and others is scale. This is important for companies like Walmart. Kroger and Albertsons are trying to merge but are held up by antitrust issues. With the new administration, there might be more consolidation. Larger companies will benefit, invest in technology, and make it harder for smaller players to keep up.