Eduardo Torrealba: Building the Honda Civic of CT Scanning

Where I thought we could start is to lay out the point A to Z. Where are we today, and what's the vision? Then we can unpack each of the individual pieces. So first, what are you building at Lumafield, and why is it so important today?

That's right. In the A to Z analogy, we might be on step B, so the very, very beginning of Lumafield. We've been building a company a little bit under the radar for a few years now, and I'm really excited to talk a little bit more about the long-term vision, where we're going and why we started the company with you today.

Lumafield was founded back in 2019 by myself and a group of co-founders. We really started this company because we want to build the organization that can take humanity from humans in the loop for design and manufacturing to humans out of the loop. It's a pretty big vision. It's not a thing that we talk about a lot at Lumafield, but that's where we're going. Step Z is: how do you have a system where a human being can say, "I want to make a new medical device," or, "I want to make a drug delivery system," or, "I want to make a new kind of battery," and then have a computer design that thing, manufacture it and ship it to the end user?

We believe that the things that hold us back from getting there are either manufacturing problems or design problems. So at Lumafield, we're building the world's best data of manufactured objects, tying that back to the design intent of those objects, and creating a system where manufacturers can actually close that loop for the very first time. That's where we're going. It's a big vision, and not the sort of thing that maybe you hear every single day. We've gone about it in a very, very specific way.

In terms of where you are today, starting with industrial CT, why was that the entry point, the wedge that made sense? And how does that inform the vision going forward?

Back in 2019, when myself and Andreas Bastian were talking about co-founding the company, we were talking a lot about this idea that problems in manufacturing and with designed objects come either from design issues or manufacturing issues. It's one of those two things, if you boil it down a lot. I'd been working at Formlabs on 3D printing, which is really on the how-do-you-make-something end of the spectrum. Andreas was working at Autodesk on generative design for Autodesk Fusion 360. We both knew that there were some gaps that you really couldn't solve independently. You couldn't perfect design without understanding manufacturing, and you couldn't get manufacturing right without understanding the design intent. So we wanted to link those two things together.

In 2019, you were starting to see the beginning of the takeoff of transformer-based artificial intelligence. OpenAI was out there. ChatGPT hadn't launched. There were a lot of really interesting things you could do with neural networks. We thought, if this takes off the way we think it's going to, these AI models will ultimately need really high-quality data to be able to make these decisions. An AI model is not going to be able to design something if it doesn't understand all of the world's design knowledge, and that's not information that's available on the open internet. It's not going to be able to manufacture something if it doesn't understand all of the world's manufacturing knowledge, another thing that's not available on the open internet. So we wanted to build a company that could be the source of truth of data for all of manufacturing.

We're not going to get there by saying that's step one. You can't go from zero to Z, at least not in this world of manufacturing. So we wanted to build something so valuable and so transformative for our customers that it was undeniable they had to buy our technology to really revolutionize the way they were working. We looked at the landscape of all the things you could do in manufacturing, and we said the best-quality data you could possibly generate in manufacturing is CT scans. So how do we make CT scanning a 21st-century technology, instead of having it be based in the same methods and the same processes that have been there since the '60s and '70s, when the technology was first birthed?

That was the idea. How do we create this data that's transformative, that people see and say, "I have to have that"? We make it affordable and easy to use. We embed it into your everyday process as an engineer, as a manufacturing organization, and then use that to create the nucleation point for all of the other design and manufacturing data you could possibly generate as a company.

That's a really good segue into the underlying technology itself. CT scans have been around for a long time, and even the underlying science is pretty old. I've heard you say that some CT scanners from 2005 can be very similar to ones today, but there is also a cutting edge getting to deeper and deeper levels of specificity. How do you think about the technological progress of this scanning discipline, and how are you pushing it forward?

I think CT scanning has mostly been pursued as a scientific instrument. It's viewed in the same way that a spectrometer might be viewed: the kind of thing that you need a really highly trained expert to operate, and not just to operate, but to actually interpret the results. In most quality assurance labs that have a CT scanner, there's this idea of level one, level two and level three trained technicians, who have maybe the equivalent of a master's or even a PhD in training, in order to understand the data and say, "Yes, this part is good, because I ran the scan correctly and I interpreted the data correctly."

That is a huge barrier. It's very, very hard to scale a system that requires that level of expertise. But we know that most people who need the data don't actually need that level. Not everything is an implant or a jet turbine engine blade. There are a lot of things, your phone, your watch, batteries, medical devices that are maybe consumables, that are important if they fail, but it's not quite the same as if a jet engine explodes in flight. So can we lower the barrier to using the technology, to make it more accessible, to make it a thing that people have as a go-to?

We really looked at the entire thing. There's the actual scanner itself. There's the fundamental technology that acquires the images. And then there's all the electronics, the compute, the software. We decided we wanted to be vertically integrated. We wanted to design our own scanners. We wanted to build our own X-ray detection technology. We wanted to run our own factory. We wanted to build our own software from scratch. We really don't buy very many things at all, and especially no software comes off the shelf. This is totally custom, vertically integrated end to end. It's a very, very unique approach to take in this industry, and it's allowed us to make innovations happen that are just not possible for other people in that space.

Beyond the skill base, there's also a cost element. Industrial CT scanners cost hundreds of thousands of dollars, if not millions of dollars. Even large organizations that have them probably have very few, and then there's a massive long tail of people who don't have them at all. If I get the names right, there are companies like Nikon Metrology and Zeiss. Are they serving the higher end while you enable the long tail? How do you fit into that status quo?

Totally. I have nothing but respect for Zeiss, Nikon and North Star Imaging. There are some really great companies that build what I think of as the cutting edge that humanity is capable of in this space. They have amazing X-ray detection technology and phenomenal X-ray sources. They're these really exquisite systems. But you don't need a Ferrari to drive to the corner to get a gallon of milk. So we're trying to build the Honda Civic of CT scanning, which hasn't existed before.

One of the sneaky things that makes CT scanning so expensive is that it's actually not just the upfront cost. If you wanted to go buy an airplane, maybe you could pull together the cash to buy an airplane. Then you've got hangar fees and maintenance and service and inspection, and if you don't get your airplane serviced every year, it's basically worthless, because who knows if it's going to work or not. CT scanners are basically the same. You have to have ongoing maintenance and ongoing support. You have to pay this money every single year to keep the thing running and maintained, because if you end up not using it anymore and you want to sell it, and you don't have these maintenance records, it's basically worthless. So your $2 million in capex goes to zero immediately.

There are some real challenges here. Your total cost of ownership actually ends up being over $100,000 a year, on top of maybe $1 million or $2 million of capex, plus the salary of the person operating it. So even if you're a manufacturing organization with a couple million bucks lying around, you might not have another $250,000 a year in your budget to run the system, because it requires so many different pieces.

We get rid of all that. We give you the hardware. We give you the software. We maintain the systems ourselves. We provide the warranty, the training, everything you need. It's a one-stop shop that starts at $75,000 a year. You don't have to pay up front for the scanner, you get more than you would ever get from any other organization for a fraction of the cost, and any engineer at your company can operate the system. We even have people who just have interns, or technicians with lots of other jobs, walk up and run scans on our system. So we've really, really lowered the barrier, not just to access the technology and acquire it, but to operate it and to actually get real information out of it.

I've heard stories of people who couldn't afford it before literally sawing open their products to figure out what's going on inside. When you think about why now, is it that there's always been a why now, and people have always wanted this capability but faced these limitations? Or is there something unique, like we're manufacturing more things, or quality control is more important than ever?

I should probably have some story about onshoring and global trade routes and supply chains, and all that stuff is true. But ultimately, we never stopped making things in the United States. We never stopped making things in countries with expensive labor. Those are usually higher-cost items that have higher quality control expectations, so you need skilled labor. Maybe there's a story about the generational shift happening, where a lot of the quality engineers in the United States are aging out. You hear that in a bunch of places in manufacturing.

But at the end of the day, this is a process that's been the same for a long time. You make something, and you make enough of it. It's actually the amount of stuff you've made that makes you better. It's not the number of years you've been in the industry. It's the amount of experience you have from making things, and that takes a lot of time in the world we're in now. For an organization making a water meter or a battery or a medical device, a new grad just hasn't had the experience of making enough of those things. So you end up with people who have been in their careers for 30 or 40 years, and they're the only ones who really know all the intricacies of how to design and manufacture these things.

There's no big change there. There's always been the question of why this thing isn't working. Factories are just an engine of entropy. If you have a factory working perfectly and you walk away from it, it's going to break in some amount of time. Maybe that's weeks. Maybe that's months. Maybe that's hours. But you need these people who are tendering manufacturing. So if we want to go faster, if we want to do more, if we want to grow, and we do want to grow our economy here in the United States, we need to be able to do more of this. We need more efficiency. We need more excellence.

So I think there's always an appetite for that, which I think is the reason you see a large number of publicly traded companies, Fortune 500s, buying our systems. Of the customers we have, 85% have never had a CT scanner before. They're big names. I can't say all of them in a podcast like this, but some of their logos are on our website. These are organizations that are cutting edge in their field, and they haven't been able to justify the cost of a CT scanner for one of their 100 facilities. Now we're unlocking that for them.

In many ways, it's just a different approach to the problem of quality control, information and data than what that industry has seen before. That's the real game changer. We came in and said: what does modern technology allow us to do? What does a combined hardware-software system allow us to accomplish? Most of the companies in this space are giant multinationals. They're not very good at building software. We're really good at building software at Lumafield.

Companies that stay massive for a long time don't stay massive because they're lumbering and sleepy. They stay massive because they're good at what they do. People would have loved to have this in the '60s, '70s and '80s, and the problem has always been there. You're taking that cutting edge and bringing it down so everybody can have it. That's a good segue into the product. The buckets I think about are hardware, software, and AI as a key component. On the scanner side you've got Triton and Neptune, on the software side Voyager, and on the AI side Atlas. There are a lot of adventurous maritime themes in there. Tell us how you think about those things, and maybe we start with the scanners and work our way down. How do they differentiate?

I would actually start with Atlas first. The long-term vision here is that you have a system that has seen a lot of manufactured objects. What if your company could hire the best manufacturing engineer in the world to come into your facility and tell you all the ways you can improve production?

Like building the brain.

Right. It's this brain that has been exposed to everything your company has ever done. We certainly don't take a competitor's data and then give it to someone else. We have two companies that make running shoes, and we're not mixing those. But you could put all of your company's data into Atlas, all of your experience, that brain of, "We're running into this problem bringing up a new product in Malaysia." "Okay, we actually had that same problem in Vietnam." You can translate it over directly, and it doesn't have to be that one engineer as the bottleneck between those two things. How do we build the corpus of manufacturing knowledge for your organization that allows you to have the best manufacturing engineer in the world on staff at all times, 24/7, globally? That's the vision of what Atlas will become. It's not that today. We certainly don't market it or position it that way. It goes back to the vision of what we're trying to enable.

But we believe that Atlas needs to be fed information and data, and that's where we zoom back to the hardware. We wanted to build a very differentiated hardware product that was easy to use, easy to make and affordable, with the hallmarks of the kind of go-to systems that just get used and abused in manufacturing facilities, and that was not a scientific instrument. So we started with: what are the pieces of a CT scanner that fail most frequently? What are the pieces that have to be replaced? What are the pieces that have to be maintained? Where is the cost per equivalent performance outside of what we think it could be if we re-engineered that product? What are the things you traditionally optimize for in a CT scanner that maybe a manufacturing environment doesn't care about as much?

We did a lot of customer research and talked to a bunch of people. I spend about one day a week with a customer somewhere in the United States right now, so I see a lot of these facilities, and I've been doing that for a long time. So we designed the hardware to be the kind of thing you would put into a quality lab in a factory. It's going to live next to equipment from Keyence or from Instron, these big scientific instrument and quality assurance companies that sell massive numbers of their units because they provide absolutely vital information to their customers. They do billions of dollars in revenue off of those product lines. That's what Neptune is. Neptune is the CT scanner for the quality lab in a factory. It's never been built before.

Then we view Triton as the next step. The core of a Triton is actually a Neptune with an automated part exchange system on top of it, so you can put it into your line. You can scan everything you're making, depending on the quantities of things you're making, and load and unload automatically. A lot of factories' quality assurance processes involve some level of batch sampling. A lot of times you care less about whether an individual item is good, and more about the trend line of all the things you're making.

Let's say you're bottling beer. Every single individual beer can doesn't have to be checked. But if you checked every hundredth or thousandth or millionth, you could actually see a trend: "Oh, actually, we're going out of nominal here. We know that this production line is starting to have a problem, and we want to preventatively stop that from happening." That's what Triton is for. It's a sampling system. For some things it'll be 100%, for some things it'll be 10%. It'll provide you enough data to control that attribute that's trying to get your factory alive. So that's Triton, a larger-volume data processing endpoint for your factory.

Then Voyager is the middle point. Voyager is the place the data shows up. It's the place you're able to examine individual scans but also aggregated scan data. It's the place you interact with Atlas. It's the focal point for the entire ecosystem. Because it's cloud-based, anybody in your entire value stream can access it if you give them that permission. So if you're making a complicated optics assembly, with parts coming from three different countries and the assembly happening in your facility, you scan it and say, "Oh, this is out of alignment. This part here and this part here are actually out of spec. We can see that through this imaging." The phrase our customers use is "smoking gun." They can send somebody a scan and say, "Hey, this part's off," and somebody on the other side of the planet can look at it, interact with that software in real time, change the process and fix that problem very quickly. It happens all the time.

On the scanner side, part of lowering the skills required is the auto-scanning capability, where you don't need to know exactly how to place the part, and it auto-adjusts. Is the ideal to completely remove any skill from the scanning process, so that literally interns can do it?

That's right. Artificial intelligence is another piece we integrate across that entire stack. Atlas is 100% AI, but Neptune is not 0% AI. There's actually a lot of software that runs on Neptune that's been trained using artificial intelligence to make sure the auto-scan settings are really high quality. Someone puts an object in, answers a couple of quick questions on the touchscreen and presses go, and it optimizes all of the system settings: the X-ray source, the X-ray detector, scan time, everything that goes into it. It's a multivariable, nonlinear equation you have to optimize to get a good-quality scan, and it turns out computers are really good at that if you train them.

Sure.

Humans are okay at it, depending on the human. I'm probably worse at it than most people. But we make it so that if you can use a microwave, you can probably use a Neptune. Depending on the settings on the microwave, that might actually be a little harder than Neptune. But we give people that next layer down. If you want to go into manual scan settings, dig all the way in and adjust everything yourself, that's available to you too. We're not going to stop experts from being experts, or people from becoming experts, but we're not going to require that as step one. I think that's the big shift. I'm sure the cameras filming us right now are probably not all running on auto mode. The film crew here knows what they're doing. But people shouldn't have to be experts to use a camera. You should have the auto mode available. That's what we've created for CT scanning for the very first time. It never existed before us.

Then you take a step down into the software side, Voyager. If I understand correctly, besides all the scan data from your scanners, customers can bring in scans from other scanners, 3D scanners or whatever it is, and bring everything together in one place. In your customers' data, how much comes from external scanners, versus all of it coming through the Lumafield value chain?

It's probably less than half a percent of the data coming from third-party stuff. We've built those integrations and that upload capability. We're pretty big believers in customer-driven product development at Lumafield. We install systems, we get people using our software, and we go visit them in, say, a die-casting facility for automotive. They have a bunch of 3D scanners sitting there, and they tell me, "I love your scanner. Your software is amazing. I really wish it was available for this product." And I'm like, "We could do that." Our software team runs a two-week sprint and releases the ability to upload other types of 3D data. Boom, that customer is uploading their data into our system.

Our sales team still focuses on getting Neptunes and Tritons in the field, and we think of the other software capabilities as places to bring that data. We talk about training that manufacturing engineer that knows everything about manufacturing, and that's ultimately going to have more sources of data than just the products we build ourselves. But we think that if you want to capture all the data from a factory, there has to be some compelling reason that you're the endpoint, and not the API feeding into the endpoint. We believe that's the unique quality of data we bring to the table with CT scanning. We think that's going to be the thing that creates the gravitational pull for other types of data, whether that's other CT scans, which is a pretty small market we're not too worried about, or 3D scans. 2D imagery is something our customers are asking us to add: the ability to just import 2D images from the cameras in their facility. They want all of that to live alongside a CT scan of the object. There's no end to the kind of data we could ingest to train that manufacturing engineering artificial intelligence. But we want to be the ones that train it. We don't want to be the ones that feed into somebody else's data.

It also speaks to accessibility. Most of these people don't have a ton of other ways to get those sources of data, so you've made that accessible for the first time. And with the value chain truly vertically integrated, people are thinking, why would I pull in all these random systems when I can have one sleek process?

That's right. We have a lot of customers who have brought us in to replace multiple other steps they do internally in their quality assurance program. They might have four or five different measurements they take, where they measure something, then cut something in half and measure something else, then cut it another way and measure that. We can eliminate all that. We can do what we call software-defined inspection. It's this idea that you do one scan, one time, and you have all of the data you could possibly need to ask questions of. Then you can use Voyager to run that analysis at scale across all the information you've ever had.

So if you're scanning every single thing you make in a factory, or 10% of everything you make, and you run into a quality assurance problem, and you didn't inspect that exact thing in the original batch, you can just create that analysis recipe, run it on your entire data set, and chart it statistically. You can say, "Okay, the problem actually only showed up in this range of time. We don't need to issue a full recall. We just need to go find these 37 items that are tied back to serial numbers in the system." So you scan once, and then you apply every single kind of measurement you want.

That isn't true of traditional measurement techniques. If you take a picture and you don't have the right angle on it, you never know what was there. You run something through a coordinate measuring machine, you run your touch probe, and you hit all the points you care about. If you don't hit the point where the problem is, you're never going to know you had a problem there. With 3D scans, you're going to get the surface, and you're not going to get the inside. We give you all of that information in one go, down to the voxel level, and you can get whatever information you want out of it. It really changes the paradigm of the way people think about quality assurance and asking questions about their manufacturing processes.

You have a lot of different types of customers: automotive, medical devices, electronics. Every device has a different shape and structure and different things the customer cares most about. When you're trying to build as much automation into the system as possible, how do you adjust for so many different end cases?

We really apply two approaches. The first is what I said earlier, customer-driven product development. We had paying customers while we were in stealth mode giving us feedback on the quality of the scans, to fine-tune what Neptune needed to be. We wouldn't build a feature without showing it to someone already paying us money, and they would tell us, "That's great," or, "I don't need that at all." I think that's really focused what we build.

The other is a real 80-20 principle: we want to build the 20% of features that cover 80% of the use cases. You're right that a medical device, a running shoe and a die-cast part for automotive are wildly different. But you might care a lot about porosity in all of those things. Are there any voids in the plastic or the foam or the metal? You can build a porosity toolkit. You can build an inclusions toolkit: is there anything in there that shouldn't be in there? That's generalizable across industries.

In the two years since we launched, we've never actually built an industry-specific workflow or technique until just now, when we launched the battery analysis capability. That one is being driven by a ton of customers who are making batteries and inspecting problems in batteries, and saying, "We really want to be able to run these specific measurements." Again, they work across enough different kinds of batteries that we can build a generalizable toolkit. We don't build super custom one-off solutions for customers. We're not a consulting business. We're building the thing that changes the game for as many customers and as many industries as possible. That's the approach we take to product development.

Before we move on to AI, I want to dig into the skills component. Part of what's compelling is that you're removing the necessity for this skill, though not eliminating the experts. I heard you tell a story about visiting a customer who had a room with four people whose only job was taking stuff apart to figure out if there were any issues. Do those people entirely go away, or do they go do something else? Technology often creates new skills, like the prompt engineer that didn't really exist until LLMs got popular. How do you think about the skills or capabilities that get created by people using your products?

My hot take is that I don't really think about that. The labor market is constantly shifting. It's moving faster than I'm going to be able to wrap my head around, and my job is not to design workforce training programs. That's a state-level thing to think about, or something for a very large organization. My job at Lumafield is to build the best technology I can to solve pressing problems that the best companies in the world need solved. If that involves upskilling labor, great. We're happy to build training programs, go in there and teach people how to use our system.

But I actually think it's more important that we think about what's the best thing for humanity in the long run. In my mind, there's one dividing line. It's not Industry 1.0 or 2.0, or I don't even know what number we're on now. Probably six or seven. I try to avoid the conferences. I think there's just a line, and it's humans in the loop versus humans out of the loop. There are a lot of things in factories that are automated that used to involve a human being, and now they don't. They involve a giant, massive press. They involve a computer-controlled milling machine, where there used to be a person with a bunch of dials. Now a computer does that. That job is eliminated, but the quality of those components is significantly better.

So our goal at Lumafield is to move humanity from humans in the loop on both design and manufacturing to humans out of the loop. We really want you to be able to walk up to a system and say, "I need a new medical device to deliver this drug. Here are the fluid properties of this drug." We want a system that can design that for you, then know where to manufacture it, manufacture it, put it in boxes and ship it to the end users. We can close those loops and get it right the first time. That's transformative. In the meantime, humans are still involved, so we'll deal with them.

But my goal is not to build workforce training programs or to build a brand-new skill set. If you ask people who work in LLMs what they think about the elimination of jobs and new job creation, the popular thing to say is, "Oh, it creates new opportunities." That might be true. That might not be true. I don't really know. But I know that the economy has grown. It's gotten better. There are more opportunities for humanity. I think as we automate more things and put those in the hands of computers, we create more opportunity for human flourishing. That's what we're doing at Lumafield.

That's reflected in how you flipped my framework. Like a lot of people, I thought about your product in terms of what I can really touch, the scanners, then what I can kind of touch, the software I can click around in, and then the AI, which is more amorphous. You flipped it around and said you think about Atlas first, and then about how to take the physical world and bring the data that's nowhere in the system to build some of this automation. It's not a question of how you rearrange the physical world. It's how you extract the data you need to do this more effectively.

As I said before, an engineer's ability to make the right decisions is not a function of how many years they've been doing a job. It's how many times they've done something. If you've done something 100,000 times, you know all the different failure modes. We really don't want to reinvent injection molding or die casting or blow molding or any of these things. These technologies work really, really well. I think that, honestly, is the downfall of 3D printing. There's been this desire to replace incredibly high-performance, very economical solutions for rearranging the physical world into the products we care about.

If we can bring the digital capabilities of the software and AI we're building and attach them to the best manufacturing systems on the planet, that's where the power comes from, not from building our own factory. You're never going to build a factory that actually covers the range of things humanity needs to produce. It's way too much stuff, and there are people who are really good at that. How do we give them superpowers? How do we bring the manufacturing engineering level you get at an Apple or a Stryker or a Lockheed Martin to every company on the planet? That's what we want to do.

Let's talk about the different pieces of the AI equation. One is helping each customer build its own specialized brain from its own data. General-purpose models that cover everything from Roman history to U.S. economics are bound to have hallucinations, and I'm sure there's some training and fine-tuning on the exact data you're creating. At the same time, I've heard you say you're not trying to be an AI model company spending billions of dollars custom-training your own models. Where do you draw the line?

It's no secret that we're using different off-the-shelf LLMs. Atlas references a few different ones, and we're constantly evaluating at what point in the conversation we should be grabbing different kinds of LLMs. It's essentially an agent. There's nothing crazy about that approach. Then we have our own data store of information that we've fed into Atlas, and then our customers' data. So there's where you pick the right LLM for different points in the conversation, what data stores you're referencing to bring up the answer to the question, and what data from the customer you're bringing to the overall conversation. We're training these things and improving them.

But today, people use Atlas to ask questions about CT scanning. The most common conversational thread is, "I've scanned a part, and I want to be able to measure a rotational slice plane through this thing. What do I need to do in Voyager?" And the answer is all the different steps you need to go through and the buttons you need to click. It's basically like tooltips for running a CT scanner and analyzing the data. It's like having a really great engineer who knows how to use the system sitting next to you and guiding you through it. It's not the co-pilot for all manufacturing today. That's where it's going. That's the future of it.

I think as we grow in that direction, there will be a world where we have to do more custom training and build more guardrails around that system. But today, we're explicitly not telling our customers to go ask it how to manufacture their device. You can do that, but there's no guarantee it's going to be perfect. Building toward that system requires a lot more energy and data than we have today. But I think we're going to be the ones that get there first because of the quantity of data we're producing.

When you look at the use cases of AI, there's data acquisition, like auto-scanning in the scanners, using AI to make it easier to get the data in. There's reconstruction, where the image is much higher quality. The thing I find most compelling is what you hinted at: a future state where Atlas helps customers make go or no-go decisions. Walk me through how you think about that. If it's not today, what would it actually look like for these systems to say, "We've looked at this, there's an issue, let us go do the things we should to make it better"?

There's that long-term A to Z vision, which is no humans in the loop, a bunch of robots and AI making everything for humanity, orbital space factories running on a Lumafield operating system. Totally crazy science fiction stuff. I think it's going to happen in my lifetime.

Today, the decision we made that I think is actually fairly unique is that we wanted to build products that would provide undeniable value on day one, and that would drive economic activity and sales of our products into our customers. The thing you see with a lot of people who approach this space is they say, "We're going to build the total solution for everything all at once." You end up hitting the entirety of the long tail of all manufactured objects immediately, and it just falls apart. This is a problem in CAD, when people say, "I'm building a new CAD tool." That's great, but until you have sheet metal, it doesn't matter. Until you can do all the things an organization would need to do, it doesn't work. You can't have 80% complete CAD tools. You need 100% complete CAD tools. So we don't want to build the 80% complete manufacturing automation system. We want to build a very specific, valuable piece that's a building block to keep going from there.

In the long term, it's not just CT scanners. It's a factory with machines that are making things, and all the inputs. Let's take injection molding, for example. The exact material you're using, the temperatures you're setting your injection molding press to, the pressure, the flow rate, the temperature of all your cooling liquids, the dwell time, everything that goes into creating that part is feeding into Lumafield. After you make the part, it gets spit out and runs through some sort of inspection process. It's Lumafield hardware. That data then goes back into the Lumafield operating system for your factory.

Then we're saying, "Okay, here are all the inputs. Here are all the outputs. Where are the differences? How do we need to attune those different parameters based on everything we've seen in the past?" If a dimensional measurement is drifting out of where we want it to be, or you're getting some warpage, or some sink, or some inclusions and voids: from everything we've ever scanned, what do we know fixes that problem? Let's turn that dial. Let's tune that thing. Let's change the dwell time for your part. Let's change the flow rate for your cooling liquid. Then we'll see how that works, and we'll close that loop continuously. The parts will be spit out, and they'll be good, and you'll be able to reach that point much faster than you can today.

Over time, we'll be able to say, "Okay, we've scanned everything your company has ever made. We've looked at it all and tuned all the parameters. Here is a design for something. Is this going to work?" We can actually run that and ask, what's this most similar to? Where are the problems? We can predict, "Okay, this line is going to need a longer dwell time than normal, because there's a thicker section in this part than it's typically used to dealing with." We can come up with a preemptive manufacturing process to tune all the parameters on a production facility, an injection molding press. You could get first-time-right manufacturing. It's never been done before, and I think we're the first people actually building the tools that can make that possible.

You mentioned that on the AI side there are a few off-the-shelf LLMs you can use, but all of your software is custom. You also have a manufacturing facility we're going to get to show off in this video. Tell me a little bit about your in-house manufacturing. Are you building a lot of your own custom components? Are there other partners you work with?

As I said before, we were thinking about what's the best manufacturing data you can produce. It's CT scanning. Okay, so how do you reinvent a CT scanner for the 21st century? We really looked at all the components that are there and how they're typically designed. Usually CT scanners are more like kit cars. They bring together different components from different suppliers, put those together, and then wrap it all in a big lead box so nobody gets cancer. We said, okay, what can we do to design all these pieces to make them less expensive? How can we use software to make control better? You don't need giant granite blocks if you have well-calibrated motion systems. So we approached it from that perspective.

As we were getting to the end of that design, the norm would have been to go to a contract manufacturer, give them your design and start building things. We've done this enough in our careers building hardware to know that the on-ramp from a signed contract to having a good unit you can ship to a customer can be six to 12 months with most CMs. We thought we could do that faster, so we decided to build our own factory, in 2020 and 2021, during COVID, when it was hard just to get into a facility. Like I said before, it's not the time, it's the number of things you've built. At this point, we're building more CT scanners of our Neptune design than anybody else in the world is building of any single model of CT scanner. So we have way more reps than anybody else, and we wanted that to be our own competitive advantage.

So we decided to vertically integrate. We found a facility and signed a lease, and I think it was 70 days later that we shipped a unit to a customer. Empty box to a fully built-out factory and a fully built, production-quality unit in the hands of a customer. I've never seen anything happen that quickly in hardware. That factory continues to produce substantial volume for us.

I get questions like, "When are you going to move to a CM?" I basically say, probably never. We're achieving 70%-plus gross margin with the way we're operating the business right now. How much better does that really need to be? If I need to get to 74%, that'd be great, but there are things I can do in software to increase my margins. I have a lot of control over my entire supply chain. We really only buy one thing from traditional vendors in the space, and that's X-ray sources, because they're good at that, those work, and there are a lot of security applications, so we can have a supply chain for them that's domestic and differentiated and doesn't depend on the Chinese government allowing us to get things. Everything else, our motion systems and the X-ray detection technology we invented ourselves, comes from a bespoke supply chain of more commodity vendors, where we can swap things out if we need to. That's pretty unusual for the space we're in.

Anything you're willing to share about pricing? I think it starts at $75,000 for the scanner, and there's also the software and other capabilities people can add. How do you think about pricing, and how do you achieve those margins when you're building your own boxes?

We want to create a lot more value for our customers than we capture, because I believe the more value you create for the customer, the stronger the relationship is and the longer it's going to last. We really don't have a lot of interest in nickel-and-diming our customers. But I also think businesses need to make a profit, be efficient and do well. So we really looked at how we provide a ton of value, and then how we make ourselves as efficient as possible.

We price the system based on some of the value we're creating for the customer, not just on what it costs us to build the system. I think a lot of the traditional hardware world is built on multipliers. Because we run on a recurring revenue business model, and because the system is half in your factory and half in the cloud, it's a little bit squishier than that. We build that value in a lot of different places. Every time you run a reconstruction, we spin up cloud resources and run GPUs, and there are costs associated with that. So we want to align the way our customers pay us with the way we're providing value. We want to bring new software features to market very quickly. We don't want to push an annual update. We want to push an update every two weeks, which is what we do in our software organization. So we move faster and iterate more quickly.

The idea around pricing is that if we're working with a customer that's building, say, the beer cans I talked about earlier, and they have Tritons across 20 or 30 locations globally, the pricing model they need might be different from a small or medium-sized business that's getting a couple of Neptunes for two facilities in the United States. So we want to price based on the value we're creating for the customer, make sure we're giving them a lot of value, and also make sure we're capturing some of it.

I can tell you I have systems where the contract is $75,000 a year, and I have a system that's $300,000 a year. The main difference there is the software, the analysis and the value we're providing. There are some hardware shifts as well, but there's not a huge degree of difference in what we build on our side. It's the same way that with a lot of electric vehicles, or even a lot of computer processors, there are software capabilities you can pay for that increase the performance. If you don't need that, you can dial it back and get a less expensive version of the thing, which makes it accessible to a lot more people. That's the way we think about pricing at Lumafield.

As you balance value creation and value capture, are there any North Star metrics you'd point to, or context you'd share, on the number of devices sold, revenue or customers?

We launched the company publicly in April of '22. By that summer, we were in the seven figures of annual recurring revenue, and we've tripled the company every year since then. So we're past eight figures of ARR at this point. In the past I'd said we were somewhere between 10 and 100 scanners. I'll say we're between 100 and 1,000 scanners. I like to stick to orders of magnitude. But this is a company growing at the speed of an enterprise SaaS company. We make margins that are best in class. We close deals really quickly. And we have a factory bolted onto the thing. Our customers have a lot of momentum. They don't return the machines. They love them, they use them, and they renew their contracts.

The business metrics are really world-class at Lumafield, and that's something I'm really proud of. But it really pales in comparison to how excited I get walking through a facility that a major CPG brand or a medical device brand is running, and hearing them talk about the transformative impact of our technology. Companies that are moving from non-sustainable plastics to sustainable materials are going to retool their entire manufacturing supply chain globally to do it, and Neptune is making it possible for them to move faster. We have companies trying to move the world from non-renewables to renewable energy products, with batteries or the electric vehicles they're making, and they have critical problems in their production that Neptune is solving for them today.

The same thing is true with medical devices. How do you get a medical device to market to deliver a specific drug? If you reformulate something that goes in an inhaler, you might need to change that inhaler. The speed of iteration there is directly coupled between the drug and the delivery mechanism, and the people who make the drugs and the people who make the delivery mechanism are not usually in the same facility. Neptune allows those people to share data, make those decisions faster and get that life-saving care to a patient more quickly. So I love the fact that we're making money. I love the fact that we're growing quickly. But I love the fact that we're really changing the lives of our customers and their customers in a really tangible, meaningful way.

Earlier you said that on the A to Z journey you're on step B. I think it's fair to say you're on step C or D. I want to end with those future steps. We've laid out the future state of what you can accomplish: removing humans as much as possible and building intelligent brains, where you have the very best manufacturing engineer in the world in your shop. What are the next few milestones over the next couple of years, and what's most important?

We think about the company in three chapters. Chapter one is autonomous data capture. Triton is maybe the beginning of the end of chapter one. How do you put a system in a facility, walk away from it, and have it capture all the data you would ever need for the other chapters?

Chapter two is what we call software-defined inspection. We've started to do a little bit of chapter two with a few of our customers. That's what I laid out around the idea that you've captured all this data, you have some sort of trend line you're looking at of the information and your production capabilities, and you're seeing something that's maybe trending out of line. SDI is this predictive thing where you can say, "Hey, seeing that the diameter of your tube is coming out of nominal, your cooling system is probably breaking down. This is how you should change it." Closing that loop will be the end of SDI, and I think that'll be a few years out. That ability to integrate all the data coming in off of all your production equipment to actually tweak those things is where we're going over the next couple of years. Triton will make the decisions on the input factors for a die-cast press, or metal injection molding, or plastic injection molding, or composite layup, any of those sorts of things that could be automated. Triton will be the driver instead of just the data feedback symbol. That's the next few years.

Further out after that is where we get into what we like to call internally wildly autonomous manufacturing, which is that post-human world that's pretty out there. It's the sense that you're not just driving the injection molding press. You're actually designing the object being produced by that press, based on the knowledge of everything you've ever produced in that facility. It's really taking the world's best manufacturing engineer and using their knowledge to become the world's best design engineer. That's more than a few years out, though. That's where we're going.