A conversation between Evolv Technology’s Founder and Head of Advanced Technology Mike Ellenbogen and Chief Scientist Alec Rose.
Evolv Technology started as a small team with a clear mission: return confidence and peace of mind to people visiting public spaces by changing the paradigm of how security professionals can assure venues are safe from the most serious threats without compromising visitor experience. While this mission was created during ongoing and escalating terrorist threats and attacks, it was well before our current global pandemic environment. But it has never been more relevant or more prescient. According to the recent Harris Insights poll, “Consumer Sentiment – Advancements in Security Screening,” the pandemic has only served to heighten consumer, employee, and visitor expectations and safety and security awareness when it comes to work, travel, shopping, entertainment, and general interactions.
It is this core mission that has attracted innovative people like Chief Scientist Alec Rose to Evolv. According to Alec, he was “doomed from the get go” when it came to math and, later, physics. He grew up with a math teacher mother and electrical engineer father and solved “fun mental math problems” from a young age. His path to Evolv, and developing complex algorithms to solve the basic idea of keeping people safe while they do their ordinary things, seemed destined.
I sat down with Alec to understand what drives him to solve the big deep-tech challenges of the space we call the digital threshold.
Mike Ellenbogen: What brought you to Evolv?
Alec Rose: I always found math fun and interesting, but I needed a real problem to apply it to. Physics was that pathway because it’s about fitting the simplest possible model to a complex problem. I definitely see myself as a physicist—I love it and am always looking for a new piece of the puzzle to learn and new tools to apply. I studied physics in college. From there I got my electrical engineering PhD degree at Duke, although even then I did everything I could to take all my courses out of the physics department to stay close to math and theory.
At Duke, I worked with Professor David Smith who was developing metamaterials for millimeter wave imaging. You and Anil (Evolv co-founder Anil Chitkara) had been following David and his work and started Evolv to essentially commercialize this work. Through this, I met you both. One thing led to another and I became Evolv’s Director of Advanced Development in 2013 and then Chief Scientist in 2020. The evolution of my role here has allowed me to blend my desire to distill a complex problem down to the simplest possible model, with the goal of keeping people safe in a non-intrusive way.
Mike: How are you able to blend these to address the deep-tech challenges of the digital threshold?
Alec: I’ll never forget one of the first things I learned in my college Intro to Physics class. We were presented with a seemingly complex problem: what happens when a horse gets struck by lightning? To break it down, our professor instructed us to start with the assumption that the horse is a giant metal sphere, because we know how to solve for this simplification. There’s no need to get bogged down by the microscopic details. Instead, always try to distill the problem down to its simplest form to get a tangible, actionable answer. I use this as a guiding principle every day at Evolv.
My graduate work was in electromagnetics and specifically the area of metamaterials, which is essentially a toolbox of solutions for creating artificial composites to solve different problems. If we wanted to bend light a particular way, if we wanted to make a particular antenna, there was a tool in the toolbox for that. I initially joined Evolv to be the “metamaterial specialist.” But I wanted to go beyond that because I was always driven by that horse analogy—to distill the most critical problems we are faced with in the digital threshold down to their simplest form and solve for that.
My path at Evolv expanded from metamaterials, to millimeter wave imaging, to reconstruction techniques within imaging—how do we consume and analyze reams of radar data, for example, to create a semblance of a person and the guns that a person might be carrying? From there, my focus quickly moved to the world of automated threat detection and computer vision. While much of my focus is now on algorithms, sensors, and their interface, as well as machine learning, I never stray too far from electromagnetics.
Mike: Venue and facility thresholds are the spaces where someone goes from being an outsider, an unknown, to a person who’s either a trusted employee, or a welcomed fan or patron. From your perspective, what are the core technical challenges that you’re drawn to in this threshold area? What are the real problems that have to be solved there?
Alec: I’m very interested in the role of the guard at the threshold. They’re often the first person that anybody meets when entering a venue. Not only are the guards responsible for spotting a gun or a bomb but they’re often asked general information questions. They suffer fatigue just like anybody else. It’s easy to blame them for long lines or missed threats. I want to make the process more synergistic with the guard. How do we make it easier for them to quickly and unobtrusively scan for, or monitor for, a threat? And all while reducing false alarms. If you lower the false alarm rate, guards are less stressed because they’re chasing fewer phantoms. And visitors are less stressed because fewer are getting stopped.
Mike: Why are there so many false alarms?
Alec: Unlike the electronic articles surveillance systems that most retailers use, we don’t have control over the shape, size, or materials of the things that need to be stopped from crossing the threshold. The possible space of threats is gigantic and it’s inevitable that in trying to protect against all of these possible threats, overlaps with some common items that people might carry will occur, creating a false alarm.
Mike: How do these overlaps occur?
Alec: The signature we look for on certain threats can be very similar to the signature on something quite benign. For example, the steel barrel of a rifle can look very similar to the steel pole of an umbrella. Since we want to catch all possible rifles while trying to let through all possible umbrellas, there’s going to be some overlap. You’re going to stop some people with umbrellas, to make sure you’re not letting through any rifles.
Compounding this, the materials used in threats can be similar to those used in everyday items; a similar amount of steel and a similar shape can show up in a gun as well as inside your laptop. At the same time, the venues and their customers are incredibly varied and very fractured. They don’t all have the same types of people coming through, they’re not all carrying the same type of “clutter,” such as bags, mobile phones or thermoses. The person coming to work at an office building is carrying something very different than the person entering a sports stadium. And each has a very different expectation of being stopped and searched based on the type of venue.
Mike: Drawing on your roots, how do you distill this into a solvable math problem? What is “the shape of this horse?”
Alec: We’ve moved well past the “if you have a hammer, everything looks like a nail” approach of traditional metal detectors that…detect metal. If we open up parameters, we can then consider not just how much metal, but what kind of metal? What shape is it? The extremely low frequency (ELF) electromagnetic waves at the lower end of the spectrum interact with metal objects and reveal what looks like just a blob to the untrained eye. But that blob still has color and shape. These two dimensions end up being immensely important to algorithms for separating consumer electronics from firearms.
And yes, this definitely harkens back to identifying the simplest possible model that can extract usable information from a very, very complex problem. How do we represent analog signals alongside digital data in the same rich way, in the right formats and with enough precision, so that they can be analyzed? In this case, the complex problem is the interaction between the system and all the possible combinations of metal objects that somebody could be carrying.
When someone walks through the Evolv system, we collect over half a million measurements across all of the different sensors and frequencies. How do we boil this vast amount of data down to actionable, real-time intelligence that the security guard can use to detect threats, make a visitor feel welcome, and not create false positives? We use a physics model called magnetic polarizability tensors (MPTs) that synthesize these half a million data points and dimensional data streams we’re constantly collecting, and represent them in six physically intuitive and computationally useful numbers. We can then teach a computer what these six numbers represent by giving it lots of examples. The computer can start drawing relationships between threats of interest and the clutter items that people carry. The guard can then use this “well-digested data” to have a clear profile of the person walking through the system.
Mike: Many solutions to problems work great in the lab, but not in the real world, where everything is dynamic and varied. How do you solve for the commercial environment?
Alec: It’s true, venues come in all shapes, sizes, ages, and infrastructures. But a commercial product needs to work in all of these environments, without exception. If we only focus on the cool things that we can do in a lab, we actually miss some of the more fun challenges of making something work in the real world.
For a security system to work, assuming it’s comprised of sensors and algorithms, the sensors need to be able to listen to their environment and adapt to temperature, to nearby metal or nearby electronics emitting in similar spectrums. Successful products actually build an algorithm that’s smart enough to listen to the environment and continuously adapt. Sensors and algorithms have to constantly verify their assumptions and be able to dynamically change in real time.
Using the Evolv system as an example, we made it sensitive to one part in 10,000 of our signal strength. When a system is that sensitive, it means anything that blows it around or moves sensors in the middle of the scan is going to present some interesting challenges. You run into this in wall scanning or synthetic aperture radar in drones, where you need to always know where your sensor is located, relative to whatever you’re imaging. It’s an incredibly difficult problem.
Mike: And then, of
Alec: As I said earlier, the goal is presenting that data from sensors and algorithms in an integrated, “well-digested” way to deliver something actionable in real time. Once the data is collected and stored, it needs to be analyzed for this actionable information, which is where machine learning takes over to look for patterns.
Additionally, not all sensors are built the same, or talk to each other very easily. You then need to create an orchestration layer to coordinate all of these different sensor streams in real time, and make sure they’re processed, that the sensors all turn on together, that they’re all collecting together, that none have failed.
Mike: Given your path to date from those early math problem-solving sessions, what do you hope will be the impact of your work?
Alec: I want to synthesize the actions that need to happen at the digital threshold down to a visitor experience that’s as unobtrusive and ubiquitous as it is at any store. My goal is to have sensors and scanning everywhere, but they are just part of the daily fabric keeping people safe while they do their ordinary things
Mike: That’s a goal we all hope for. Thank you, Alec.
