Verdun launches an AI camera-monitoring service with trained human operators

Verdun launched on Y Combinator Fall 2026 on 28 October. The YC launch page had 4 votes. The company is a remote video monitoring service. AI agents review every camera alarm, then escalate real threats to trained operators who coordinate a response. The founders are Sonny and Tung.

The product has four parts, Verdun says on its launch page. AI triage dismisses false alarms with the reason written down. Real threats are flagged in seconds. Trained human operators confirm the threat and decide the response. The options are a voice-down over on-site speakers, a call to listed contacts, or dispatching police or a guard. Voice-down is handled by AI voice agents and operators, warning intruders off through horn speakers at the site. The product reports camera health, incident breakdowns, response times and a full log of every agent and operator decision.

What makes the pitch different from existing products

Monitoring centres today sell on human attention, Verdun says. The problem is that human attention does not scale. Centres cannot add staff when alarms spike. They mute noisy sites and dismiss alarms in bulk. Verdun writes they do this "hoping that they didn't just wave off a true threat". Traditional video analytics then flag every person or vehicle that crosses the frame. The operator sees too many alarms and dismisses them anyway.

Verdun's claim is that the agent reads context first. It judges whether the activity is normal for that site and time of day. If a forklift drives through a yard at 2 pm, that is routine. If a person tries the cab door of a parked truck repeatedly at 2 am, that is a threat. The site's demos show both cases. One is dismissed; the other is escalated to an operator for a voice-down.

Who built it

Sonny is a former United States Marine officer and a repeat founder. He previously scaled a computer-vision object-detection company to over 100,000 cameras, the launch says. Tung holds a PhD in applied mathematics. He is a National Math Olympiad Gold Medalist. He was a senior AI engineer at LinkedIn, at VergeSense (YC S17) and at WeWork. The team is in San Francisco.

What is missing from the launch

The launch names no customer. The site shows an "OpsView" with "added people and classifications simulated". The company does not publish a false-positive rate, an average response time, or a comparison against a traditional monitoring centre. The ask on the page is for introductions. Verdun wants multi-site businesses with cameras, especially car-dealership groups. It also wants security dealers and integrators who sell monitoring.

The reader of this item is not usually in security operations. The reason to pay attention is the shape, not the vertical. An agent reads context. It writes down its reasoning. It hands the real cases to a person. This is a reusable design. It fits anywhere a stream of alarms is more than humans can read. Compliance alerts, fraud alarms, product-abuse signals, and infrastructure monitoring all look like this. Verdun is building the pattern for commercial cameras. The test for a buyer is accuracy on the buyer's own site, measured in writing. The product writes down its reasons. That makes the test easier to run than for a traditional analytics vendor.

Source

• Primary source: Verdun on Y Combinator.
• Product site: Verdun.