Interviews

Andrew Nassar, CTO of IC Realtime – Interview Series

Antoine Tardif
Add Unite.AI to your preferred sources on Google
Andrew Nassar, CTO of IC Realtime, against a light gray background.
Andrew Nassar, CTO of IC Realtime.
Key PointsBunker runs core AI analysis on a local NVIDIA Jetson gateway to cut cloud dependence…
View more
  • Bunker runs core AI analysis on a local NVIDIA Jetson gateway to cut cloud dependence and speed alerts.
  • Nassar refuses to build biometric identity features because surveillance should describe events rather than profile people.
  • Edge AI is not a privacy policy since cloud services still handle remote access and video history.

Andrew Nassar is CTO of IC Realtime, where he has helped guide the company through shifts from analog cameras to digital recording, remote monitoring, and AI-assisted video analysis. With more than 20 years in the surveillance industry, he focuses on turning advances in computer vision and edge computing into practical tools for understanding events, finding footage, and improving security operations.

IC Realtime designs video surveillance systems and software for residential and commercial environments. Its Bunker platform adds AI analysis to compatible camera systems through a dedicated NVIDIA Jetson-powered gateway, supporting contextual alerts and natural-language video search. The platform combines local analysis with cloud-connected services for remote access and video history.

You’ve helped lead IC Realtime through the transitions from analog cameras to digital recording, remote monitoring, and now AI. What convinced you that AI could fundamentally change surveillance, and what advances have made sophisticated video analysis practical at the edge?

We have worked with machine learning and computer vision on our cameras for years. The opportunity became much more useful when the system could help someone understand an event and find the relevant footage, rather than simply produce more video to review. A homeowner wants to know what deserves attention. A business wants to find an incident without spending hours searching.

Bunker brings that approach to a dedicated NVIDIA Jetson gateway at the property. Moving core analysis closer to the cameras can reduce the dependence on sending every analytical task to the cloud and improve responsiveness. The practical advance is the combination of local computing and software that turns video into contextual alerts and searchable events.

My standard is still proven value. The technology has to solve a real problem in the environment where the customer will use it. A good demonstration is the beginning of that process.

Using Bunker as an example, could you walk us through what happens between a camera capturing an event and a user receiving an alert? Which AI functions run locally, what still requires the cloud, and what data leaves the property?

At a high level, a compatible camera sends video to the Bunker Gateway, and core AI analysis happens on that gateway at the property. The platform links detected activity to video events and can send a contextual alert with a snapshot and a plain-language description. The user can then review the video in the app.

Bunker uses a hybrid architecture. Cloud connectivity supports remote access, subscription services and video history, so it would be misleading to call the platform cloud-free or to say no video leaves the property. What matters is how that data is protected when it does leave. Everything is encrypted in transit, including live video, which stays encrypted from the gateway all the way to the viewer’s screen. The infrastructure in between relays it but can’t watch it. Each gateway authenticates with its own certificate, never shared credentials, and footage stored in the cloud is encrypted.

Customers should get a clear explanation of that flow, including which features depend on the internet and what happens when the connection is down.

Security cameras operate in rain, darkness, glare, and constantly changing environments. What is the hardest real-world failure you’ve encountered, and how do you measure whether improvements reduce false alarms without increasing missed incidents?

I’ve been in this industry for more than 20 years, and camera technology has improved dramatically, especially in low light and difficult conditions. Nighttime and noisy images are now nearly as clear as daytime. Even so, rain, glare, fog and snow can still reduce an AI model’s accuracy. Better sensors don’t remove those variables.

So we chose a dynamic approach that adjusts to changing environments and times of day, rather than relying on one fixed configuration. Over the past two years of development, we’ve put thousands of hours into testing in real-world installations, not just cameras on a bench, and we’ve built the system around what those deployments taught us.

I wouldn’t judge an improvement by a lower notification count alone. The question is whether it removes irrelevant alerts while still catching the incidents that matter. Reducing noise is only valuable if the system stays useful when something important happens.

Detecting a person or vehicle is different from deciding that behavior is concerning. How do you define an anomaly, adapt that definition to each location, and prevent the system from treating harmless activity as a threat?

A detected object does not, by itself, tell you whether there is a problem. A delivery at the front entrance and activity in a restricted area can involve the same type of object but require very different responses. The starting point should be the customer’s use case and the context of that particular property.

I would distinguish an event worth reviewing from a conclusion that someone is a threat. The system should help a person understand what was observed and why it was surfaced, with access to the underlying footage. It should not encourage the user to treat every unfamiliar activity as evidence of wrongdoing.

As these systems become more capable, the challenge is to make that context useful without overstating what the AI knows. Human judgment remains important, particularly when an alert could lead to a consequential action.

Natural-language search promises to replace hours of manually reviewing footage. How does your system connect a user’s description to the right video, and how do you test for missed events, incorrect matches, or AI-generated descriptions that the footage doesn’t support?

The concept is to make video searchable in the language a person would naturally use. As activity is detected, Bunker creates time-stamped descriptions associated with video events. A user’s question is used to search that indexed record and surface relevant footage. That gives the user a starting point without manually scrubbing through the entire recording.

The footage should remain the basis for checking the result. A generated description is useful for navigation, but it should not be treated as more authoritative than what the camera actually captured.

For testing, I would separate three questions: Did the system miss a relevant event? Did it retrieve something that does not match the request? Did the description include something the footage does not support? Those are different failure modes. Evaluating only whether the search returns a plausible-looking answer would not be enough.

What trade-offs does NVIDIA Jetson hardware impose on model size, accuracy, power consumption, and the number of camera feeds you can analyze? How do you maintain reliable performance when using older third-party cameras, reaching processing capacity, or losing internet connectivity?

The design question is how to deliver useful analysis within the resources available at the property. Model complexity, the amount of video being processed and the required response time all have to fit the gateway’s operating limits. That is why the number of supported streams alone does not tell the whole performance story.

Bunker’s launch plans support up to 10 camera streams. Compatible ONVIF cameras let customers keep their existing hardware, but compatibility still needs to be verified for each installation. Profiles, firmware, image quality and the features being used all matter.

We also need to be precise about resilience. Local processing and cloud-connected services are different parts of the system, and having an edge gateway doesn’t mean every feature works normally during an internet outage. Updates are part of reliability too. Every update is signed and verified on the gateway before it installs, older vulnerable versions are rejected, and a failed update rolls back automatically. A bad update shouldn’t be what takes a property’s security system offline.

Local inference alone doesn’t answer every privacy question. Where does Bunker store footage, search queries, and AI-generated metadata, how long are they retained, and who can access them? How do you improve your models without assuming access to customers’ private footage?

I agree. “Edge AI” is not a complete privacy policy. Customers need to understand where their data goes, what is retained and who can access it, and that includes search queries and generated descriptions as well as video.

Footage and the databases behind the service are encrypted at rest. Passwords are hashed and sessions expire, so even a database breach wouldn’t give anyone working logins. The gateway itself is encrypted too, so a stolen device yields no footage and no credentials. Bunker’s launch offering included 60 days of cloud video history on Standard and 365 days on Premium, plus Video Vault for clips a customer chooses to save.

We don’t use customer footage to train our models, and we don’t create biometric templates of anyone the cameras capture. A customer using the product is not permission for us to reuse their private video.

The controversy surrounding Flock Safety has highlighted questions about who can search surveillance data and whether access is properly revoked. Although your deployment model differs, what lessons apply to your systems, and which safeguards should be mandatory rather than optional?

The backlash against Flock isn’t really about the AI. It’s about who can get in and search the data. Much of the scrutiny came down to account security, including lawmakers asking why multi-factor authentication wasn’t mandatory. When a stolen password is enough to search a surveillance system, it doesn’t matter where the model runs.

With Bunker, the only people with access to a system’s data are the users that system’s owner has granted access. We don’t share it with third parties, we don’t use customer data to train our models, and we don’t create biometric templates.

Security has to start at the device and carry through to the account. Every Bunker gateway ships with its own cryptographic identity, so there are no default passwords to change or for an attacker to look up. Firmware is verified at every boot, updates are signed and checked before they install, and logins are protected with multi-factor authentication. Beyond that, access should be removed promptly when it’s no longer needed, and sensitive activity should leave an auditable record. Trust depends on being able to explain who did what.

People captured by a camera often aren’t the people who bought it. How should AI surveillance account for the privacy of neighbors, employees, visitors, and passersby, and are there capabilities you would deliberately refuse to build?

The installation needs a defined purpose, and that purpose should guide what is captured, what is analyzed and who can see the results. Wanting to secure an entrance doesn’t mean every activity on a neighboring property should become part of a searchable record.

One line we’ve drawn is identity. Bunker doesn’t create biometric templates, and it doesn’t attach an identity to anyone it detects. It describes what happened, not who someone is. That keeps the system focused on events rather than on tracking or profiling people.

More capability doesn’t automatically mean more value. If a feature creates a serious trust problem without solving a clear customer problem, that’s a reason not to build it.

Looking ahead three to five years, what meaningful capability do you expect edge AI to unlock beyond today’s detection and search—and what technical or trust problem must be solved before it deserves widespread adoption?

Honestly, the future of AI is uncharted right now, and it’s moving faster than anything I’ve seen in my career. We’re following those changes almost daily and looking at where they can genuinely benefit the people using our platform.

The areas that matter most to us are context, understanding, and systems that keep learning and improving. The value is in helping someone understand a developing situation and decide what needs attention, instead of piecing together a string of separate notifications.

The technical challenge is to do that reliably in changing environments and within the limits of the local hardware. The trust challenge is to make the system’s limits and data practices easy to understand. People need to know what it observed, what it inferred and when a human needs to make the call.

Thank you for the great interview, readers who wish to explore professional security cameras, recorders, and AI-powered video surveillance systems should visit IC Realtime.

Antoine is a visionary leader and founding partner of Unite.AI, driven by an unwavering passion for shaping and promoting the future of AI and robotics. A serial entrepreneur, he believes that AI will be as disruptive to society as electricity, and is often caught raving about the potential of disruptive technologies and AGI.

As a futurist, he is dedicated to exploring how these innovations will shape our world. In addition, he is the founder of Securities.io, a platform focused on investing in cutting-edge technologies that are redefining the future and reshaping entire sectors.