Verkada - Your AI-Powered Enterprise Security System 🔒

CV Deep Dive

Today, we’re talking with Yunchao Gong, Head of Camera Software Engineering at Verkada.

Verkada is a fast-growing security technology company that builds cloud-based enterprise security solutions, including video security, access control, alarms, sensors, intercoms, and workplace management - all integrated into a single platform. Founded in 2016 by graduates of Stanford University and MIT, the company is best known for its commercial-use camera systems, which use AI and computer vision to enhance security whilst maintaining a privacy-sensitive approach.

Today, Verkada serves a wide range of customers, including gyms (i.e. Equinox), retail chains, school districts, hospitals, manufacturers, and state governments. The company has a large engineering organization dedicated to developing AI-powered systems for security, optimizing computer vision models for edge processing, and maintaining an extremely high-performance engineering culture.

In this conversation, Yunchao discusses Verkada’s approach to AI, the challenges of running computer vision on-device, and how the company empowers its engineering team.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Yunchao 💬

Yunchao, welcome to Cerebral Valley! First off, introduce yourself and give us a bit of background on you and your role at Verkada. 

Hi there! I’m Yunchao, and I’ve been with Verkada for about five years. At Verkada, I lead the camera software engineering team, which is responsible for building all the software around our video security cameras.

I’ve been working in the computer vision for more than a decade. I earned my PhD in computer vision 15 years ago at the University of North Carolina at Chapel Hill. Back then, computer vision wasn’t nearly as popular as it is today. After that, I was lucky to join Facebook AI Research as one of the early members, working directly with Yann LeCun for a year.

From there, I joined Snapchat, moved to LA, and became one of the company’s first 100 employees. I spent four amazing years in Venice Beach before moving into autonomous driving at Zoox, where I worked for two years before eventually joining Verkada.

How would you describe Verkada to the uninitiated enterprise or AI team?

Verkada is solving a critical problem—safety. If you look at most Bay Area startups, very few are focused on this space, which is why I think that Verkada is such an important company. What we’re building makes a real impact in the world. 

We build enterprise security solutions, including video security systems, access control, alarms, sensors, intercoms, and workplace management systems that help more than 30,000 organizations across the world protect their people and places. All of these work together on a single, easy-to-use software platform that our customers can access from anywhere - even their phone. It might seem like a no-brainer that all of these building operations tools are connected and easy to use. But the reality is, it actually had never been done before Verkada.

Who are some of Verkada’s key customers today? Who is finding the most value in the systems you’re building there?  

Verkada has customers across all sizes and shapes.When you really zoom out, any building or space that needs to manage access or be secure can be a customer,. As someone building these solutions, it’s an exciting challenge that we aren’t limited to serving just one vertical or industry.

Our customers are brands that you probably know, and maybe you’ve seen Verkada “in the wild” without realizing it. For example, at the gym.  Equinox was actually one of our early adopters. Many retail brands also use Verkada for loss prevention, ranging from the Goodwill to luxury brands like Canada Goose.

Another area we’ve seen widespread adoption is schools. Nothing is more important than the safety of children in schools, and I’m so proud that we protect so many schools across the country – even my local school district. 

Beyond that, hospitals and manufacturers are adopting Verkada as well. One great example is Dairy Farmers of America, which relies on our platform for security and operational monitoring.

At the end of the day, as long as there’s a building, there’s an opportunity for Verkada to provide security – whether it’s video security and alarms or visitor management and access control.

Walk us through Verkada’s flagship products. What would you say are some of the key innovations that have enabled Verkada to establish itself at the forefront of AI-enabled security? 

I run the camera software engineering organization, so my answer will be a bit biased towards the ways we’ve integrated AI and other frontier technologies into our camera product. If you ask me about the key innovations behind Verkada, I’d highlight three main things.

First, we are a cloud-based platform. We were not only a pioneer in cloud security cameras but also one of the largest players in this space. Being cloud-native is a major innovation for Verkada and we’ve been a key driver behind the industry moving towards the cloud.

Second, AI has been in our DNA since the company’s founding. We’ve been deeply invested in AI and computer vision from the start, continuously innovating and expanding our capabilities in this area. I’m really proud of the features that we continue to ship that ultimately help our customers to be more efficient and effective in keeping people safe.

Third, we focus heavily on user experience and product design. We are a user-driven company, committed to delivering the best possible experience for our customers. As our CEO puts it, we are building a "single pane of glass"—a unified platform where all our products, including cameras, access control, alarms, sensors, and workplace management, work seamlessly together.

Verkada has always been known for frontier innovations in ML, but how is Verkada leveraging this new wave of generative AI? 

AI is in Verkada’s DNA. It was one of the key competencies we set out to build when the company was founded.

Early on, Verkada already had a wide range of AI and computer vision features. For example, we perform powerful object detection and tracking directly on the device, analyzing more than 15 frames per second with an object detector running in real-time. For our customers, this means that they can quickly identify and track potential bad actors or persons of interest.

We also implement state-of-the-art attribute analysis, generating detailed attributes for people and vehicles to help customers conduct investigations. Our search and alerting features  leverage the latest advancements in AI to provide fast and accurate results for customers when they need to find footage of people or vehicles.

These technologies were built using state-of-the-art computer vision research, particularly convolutional neural networks (CNNs), which remain highly effective. Since the rise of large language models (LLMs), Verkada has been actively exploring new AI capabilities to ensure we remain a leader in AI-driven security solutions.

In late 2023, we introduced a groundbreaking innovation leveraging one of Open AI’s vision-language models (called CLIP) to map images and language into the same latent space using transformers. This enabled a completely new product experience, allowing customers to search for relevant images and footage using everyday language—whatever they want to type.

For example, you can enter a query like "white tesla with a dent on the bumper," and the system will return relevant footage matching that description. 

This innovation is significant because we were the first in the physical security industry to launch this capability. We have a competitive edge as we continue to iterate on and improve the product experience. For example, we shortly followed this natural language search feature with an “alerting” option so that customers could get real-time alerts for the types of queries that they were searching for. This is especially helpful in manufacturing settings where our customers not only use Verkada for security, but also for safety. You can imagine that alerts for “person not wearing a safety vest” or “person driving a forklift with headphones in” can be powerful tools for supervisors tasked with ensuring safety and compliance of their teams.

At Verkada, how do you measure the impact you’re having on security for your customers? Are you monitoring accuracy, volume or something else? 

It's a complicated process, and I don’t have just one magic number to share. We evaluate our system in two key ways—one quantitative and objective, the other subjective and customer-driven.

The quantitative approach involves tracking hundreds of different metrics to measure system effectiveness. We monitor latency, success rates, computer vision accuracy, and notification delivery speed—for example, ensuring that every notification is delivered within seconds. These metrics are displayed in a dashboard, reviewed weekly, and monitored in real-time with alerts to catch any issues.

However, metrics alone don’t always tell us whether customers find the product genuinely beneficial. That’s where the subjective approach comes in. We talk to customers regularly, engaging with both prospects and existing users, and frequently visit them on-site to see firsthand how they use the product and what feedback they have.

What truly sets Verkada apart from others in this industry is that we don’t just listen to feedback—we act on it. We take a customer-centric approach, sharing insights with our team and incorporating them directly into our product roadmap. This continuous feedback loop ensures we build what customers actually need, making our platform even stronger over time.

How are you balancing the dynamics between engineering and research in this new era of generative AI, given so much of the innovation is taking place at the research level? 

There’s no easy answer to this because research is inherently risky. Most research efforts don’t turn into successful products, but at the same time, engineering and research time is incredibly valuable and expensive. We can’t afford to invest indefinitely in research without seeing results.

So, we approach it in a few key ways. First, we avoid taking on projects that are too risky. We use our understanding of research and technical feasibility to eliminate ideas that are unlikely to work. But beyond that, there’s a large body of research that is close to production with moderate risk, and for these, we encourage people to be open-minded, build prototypes, and keep experimenting.

The only way to know if a research idea can become a successful product is to build it. Spend a couple of days, create a simple prototype, and see if it wows people. One of our best examples is our AI-powered search feature. It started as a simple prototype built by one engineer. They sent it to me, and I was immediately blown away. I spent a few hours playing with it and then called our CEO, telling him, “This is amazing, you have to check it out.” His response was, “Send me the link.” But at that point, it was just running on a developer’s local machine, so I told him, “It’s an IP address on someone’s dev box.” He said, “Bring me your laptop, I want to try it.”

This is the mindset we encourage—build fast, test ideas quickly, and see what resonates while keeping the initial investment low. Not every experiment succeeds – but when they do, they can have incredible impact. That’s why our philosophy of rapid prototyping, testing, and refining ideas makes Verkada such a fun place to work.

What has been the hardest technical challenge around building Verkada into the platform it is today? 

One very ambitious goal we set out to achieve five years ago when I joined the company was to build all computer vision on the edge. Back then, this was incredibly challenging. The chips weren’t as powerful, and every model Verkada ran was on the server, using huge models that were difficult to scale. Five years ago, this was a very ambitious goal.

As of today, we have successfully shipped almost all our models on the edge—95% of computer vision in Verkada now runs directly on the camera itself. This was made possible by three key factors:

1. Chips have become more powerful, now featuring specialized processing units for neural networks.

2. The industry and research community have evolved, continuously innovating to develop smaller, more efficient models that run effectively on chips.

3. The incredible work done by Verkada engineers, optimizing models to be small enough, efficient enough, and performant enough to run on-device.

These advancements have enabled us to overcome major technical challenges and fully realize the vision that the founders set out to achieve.

How do you see Verkada evolving over the next 6-12 months? Any specific developments that your customers should be excited about? 

I can't get into too much detail, but Verkada is continuing to double down on AI. Large vision-language models are becoming a major focus in the industry, and with the latest developments in Transformers and other AI advancements, we are now able to tackle computer vision challenges that were previously unsolvable. For example, activity recognition and other complex vision tasks are becoming more feasible.

We are actively exploring the latest research and leveraging new breakthroughs to solve hard problems that previous models couldn’t handle, bringing cutting-edge AI capabilities to our customers.

Privacy is a huge component of any security system - how are you incorporating this into Verkada’s wider systems given the additional risks with new AI models? 

Verkada's mission is to protect people and assets in a privacy-sensitive way, and privacy is at the core of what we do.

Privacy is embedded in our product design. For example, all facial recognition features are turned off by default, meaning customers must explicitly enable them and agree to specific terms before use. We also ensure that all actions are logged and audited, and give granular permissions so that only the appropriate users within an organization can access features that, for example, leverage facial recognition. 

We have also innovated several privacy focused features for our customers:

• Privacy regions, allowing customers to mask sensitive areas.

• Face blur, which blurs faces in live and historical footage unless toggled “off” by users with appropriate permissions.

• POI-only search, a privacy-respecting feature that limits user searches to designated threats or persons of interest.

When we launched AI search powered by our large vision-language model, we invested heavily in query moderation to ensure privacy protections. This system provides guardrails, preventing users from searching for terms that could be disrespectful or violate privacy standards.

How would you describe the culture at Verkada? Are you hiring, and what do you look for in prospective team members joining the Verkada? 

We hold an extremely high bar when hiring. We’re looking for intelligent, hardworking people to join the team who really care about our mission and vision. We’re doing really important work, and it’s important that our team is up for the challenge.

But with the challenge and high bar that we set, we also have a culture that rewards results. This encourages our team  to stay focused on delivering impact and outcomes, and it’s how I believe we’ve built such an incredible product so quickly and ultimately how we continue to stay at the forefront of innovation.

Conclusion

Stay up to date on the latest with Verkada, learn more about them here.

Read our past few Deep Dives below:

• Build specialized RAG agents with Contextual AI 🌐

• Remyx - Your AI Production Assistant 💡

• Qdrant's GPU-accelerated vector indexing is here 🔋🌐

• Athena is your AI-powered remote hire 🎛

• Doowii - your AI-first education platform 📖

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.

All Events | Jobs