The real bottleneck isn't detection accuracy. It's power.

"There are institutions, like county government offices, with a fixed power limit."

That's not the answer you'd expect from an AI video surveillance interview. Most people assume the hard part is spotting incidents accurately or fast enough. It turns out that's only half the story.

The other half is deployment reality: can the server draw enough power, does it run on the operating system already in place, does it talk to the existing monitoring system? None of that shows up on a spec sheet. It's exactly the set of questions Intellivix had to answer when choosing hardware for the Korea Expressway Corporation's AI CCTV conversion project, and it's why they landed on Mobilint's NPU (Neural Processing Unit, a chip purpose-built for running AI inference efficiently).

We sat down with Kang Changhee, an AI Researcher at Intellivix who leads the joint technical work with Mobilint, to hear how the project actually came together.

Kang Changhee, AI Researcher at Intellivix

A national-scale rollout, without ripping out a single camera

Intellivix has spent years working on how AI can scale safety monitoring. This project, run as part of a public AI CCTV program led by Korea's National Information Society Agency (NIA), converts roughly 4,000 channels of the Korea Expressway Corporation's highway CCTV into an AI-driven monitoring system, covering tunnels, ice-prone stretches, accident hotspots, and chronic congestion zones nationwide.

The existing camera infrastructure stays exactly as it is. What changes is what sits on top of it: AI reviews each situation first, and only flags confirmed events to a human operator.

Two AI passes before a person ever looks at the screen

Picture a single stopped car on the highway. Is it a breakdown, an accident, or just traffic? Multiply that ambiguity across hundreds of camera feeds running at once, and you can see why a simple "vehicle stopped" alert isn't useful on its own. Confirming what's really happening takes context: the vehicles around it, the road conditions, how the scene evolves.

That's the gap Intellivix built its system to close, by pairing Vision AI with a VLM (Vision-Language Model, an AI model that can reason about both images and text together).

System diagram of the Korea Expressway Corporation's AI CCTV conversion project

Vision AI, running on Intellivix's VIX2.0 engine, does the first pass: real-time detection of accidents, stalled or stopped vehicles, wrong-way driving, and road debris. The VLM then does a second pass, re-checking the scene before and after the flagged moment against the surrounding road context to confirm whether it's a genuine incident.

The hardware carrying that second, heavier step is Mobilint's ARIES-based MLA400 PCIe Card.

"Existing video analysis systems can detect objects quickly," Changhee explains, "but complex road conditions, like vehicle shadows, shifting light, weather, or ordinary traffic jams, can trip them into flagging a normal situation as an accident."

Combining Vision AI's speed with the VLM's second look cuts down those false positives and raises confidence in every flagged event. Verified incidents then route through Intellivix's Gen AMS platform, so operators review only confirmed events instead of scanning every feed themselves. The result, in Changhee's words: far fewer alarms for monitoring staff to chase, and faster, more focused response when a real incident does happen.

The Gen AMS platform, which understands video and converts it into text, refining it into reports for people (Source: Intellivix)

Why Mobilint, specifically

Designing that two-step verification pipeline is one problem. Running it across hundreds of channels, in a live government deployment, is a different one. Changhee's evaluation came down to three concrete criteria: realistic power draw, OS compatibility, and support that actually shows up when needed.

Power headroom, without adding power infrastructure

Some of the sites in this rollout, like county government facilities, operate under hard power caps. GPU-based servers draw enough power per unit that scaling up would normally mean adding supply infrastructure like a UPS*. With Mobilint's hardware, Intellivix increased the number of channels a single server could handle without adding any power capacity at all.

*What's a UPS?
Short for Uninterruptible Power Supply: a device that keeps connected equipment running on stable power through outages or voltage swings.

The only NPU with Windows OS actually verified

Most monitoring centers, this project included, run on Windows-based server environments. When Intellivix compared vendors, Mobilint was the only one whose NPU had real, verified Windows driver support, not just a claim of compatibility. "The other vendors we compared hadn't gotten that part to the verification stage," Changhee says.

The 'MLA400 PCIe Card Passive Slim,' equipped with four of Mobilint's low-power, high-performance AI accelerators, 'ARIES,' deployed in this project

Fast support when the model work got hard

Intellivix's system runs more than 10 detection models alongside VLMs at once, all tuned for a safety-critical use case. Since the VLM step needs more compute than a vision model alone, the CPU-NPU pipeline had to be tuned carefully so it responds the moment an event is flagged, with no lag.

"There were plenty of details to work out early on," Changhee says, "but Mobilint's model team responded fast every time we had a question, just like on earlier projects we'd done together. That's what got us through it smoothly. It's stable in production now, and this project is moving forward with the same kind of quick support."

Researcher Kang Changhee running the VLM system at his desk. The MLX-A1 Edge AI Box on the desk stands out.

"Because of the track record we built through earlier local-government pilots, choosing Mobilint again for a much bigger, nationwide project wasn't a hard call."

Beyond this project: Intellivix's push toward "Safety AI in Action"

Intellivix has been building its name across several fronts recently. At exhibitions in Korea and abroad, the company has shown off a four-legged patrol robot, ARGOS, and VIXA, a VLA-engine safety AI agent, under the banner "Safety AI in Action." It's also the first company in Korea to hold all five major international standard management certifications, including ISO 42001, the international standard covering AI reliability and accountability.

The vision and slogan that welcome visitors at the entrance of Intellivix's headquarters: "The Safest World with AI"

Certificates and plaques filling every corner of Intellivix's headquarters

"'Building a safer world with AI' is where we start.
We combine Vision AI with generative-AI-based VLM technology, and our mission is to protect people's lives and property through it.
For AI tied directly to safety, we believe performance alone isn't enough. Reliability and accountability have to come with it, or you don't earn the trust of customers or society."

None of this required Intellivix to rip out its existing camera network. Instead, AI now backs up the judgment already made by human operators, and that combination is set to make highways across the country safer, one rollout at a time.

Changhee's advice for other companies weighing a Mobilint partnership: "If you're processing large volumes of video in real time, Mobilint is worth a serious look. You get low-power, high-efficiency hardware that still plugs into legacy systems like existing VMS or NVR setups, so you can add AI capability without replacing your facilities wholesale."

And his one-line summary of the partnership's goal: "Intellivix's Vision-VLM precision, paired with Mobilint's efficient NPU. Together, we're building the most trustworthy AI safety net we can, for as many people as possible."

Want to see how this could apply to your own environment?

Intellivix's safety AI platform, including the Vision AI and VLM combination and the Gen AMS reporting layer covered in this story, is detailed on the Intellivix website.

→ Visit the Intellivix website

The NPU behind this deployment, Mobilint's MLA400 PCIe Card, has full specs and performance data on its product page.

→ Visit the MLA400 PCIe Card product page

This story is based on an interview with Intellivix AI Researcher Kang Changhee, conducted in early September 2026. The figure of approximately 4,000 channels refers to the scale of the project, not the processing performance of any single card or server. Details may change as the project progresses.