The surveillance technology sector is currently facing significant scrutiny, though not for positive reasons. Recent controversies include U.S. Immigration and Customs Enforcement accessing Flock’s camera network to monitor individuals, and home security company Ring facing backlash for developing features that allow police to request footage from residents' outdoor cameras. These developments have sparked a widespread discussion about safety, privacy, and the boundaries of observation.

However, controversy does not eliminate market opportunities. Advances in vision-language models have provided fresh momentum for firms creating innovative solutions to help businesses monitor their facilities. According to Matan Goldner, co-founder and CEO of video surveillance startup Conntour, ethical considerations are so crucial that his company is highly selective about its clients. This approach might seem risky for a startup only two years old, but Goldner notes that Conntour can afford this selectivity because it already serves several major government and publicly-traded organizations, including Singapore’s Central Narcotics Bureau.

“Having such significant customers allows us to maintain control over who uses our technology and for what purposes,” Goldner explained. “We can choose what we believe is both moral and, naturally, legal.” This established traction has done more than enable selectivity—it has also attracted investor attention. The startup recently secured a $7 million seed round from General Catalyst, Y Combinator, SV Angel, and Liquid 2 Ventures. Goldner mentioned the round was finalized in just 72 hours. “I scheduled about 90 meetings over eight days, and by Wednesday afternoon of the first week, we were finished,” he said.

Conntour’s cautious approach may be well-founded, particularly given the advancing power of AI tools in this field. The company’s video platform employs AI models to allow security staff to query camera feeds using everyday language, searching for specific objects, people, or situations in real-time—essentially creating a Google-like search engine for security footage. It can also autonomously monitor for threats based on predefined rules and automatically generate alerts.

Unlike traditional systems that rely on fixed parameters to identify objects or behaviors, Conntour states that its platform utilizes natural and vision language models, offering greater flexibility and ease of use. For example, a user could ask, “Find instances of someone in sneakers passing a bag in the lobby,” and the system would swiftly scan recorded or live video to provide relevant clips.

By integrating AI models, the platform allows users to ask questions about footage and receive text answers alongside the corresponding video segments, as well as automatically generate incident reports. A key selling point for Conntour is its scalability. Goldner emphasized that the platform stands out from other AI video search services because it is built to efficiently handle systems with thousands of camera feeds. In fact, he noted that Conntour’s system can monitor up to 50 camera feeds using just a single consumer-grade GPU, such as an Nvidia RTX 4090.

This efficiency is achieved by employing multiple models and logic systems, with the algorithm selecting the most suitable and least computationally intensive option for each query. The company claims its system can be deployed entirely on-premises, fully in the cloud, or in a hybrid configuration. It is designed to integrate with most existing security setups or can operate as a standalone surveillance platform.

A persistent challenge in video surveillance, however, is that effectiveness depends heavily on footage quality. Details can be lost in poorly lit areas or when using low-resolution cameras with dirty lenses. To address this, Conntour includes a confidence score with its search results; if a camera feed is of insufficient quality, the system will flag results with low confidence levels.

Looking ahead, Goldner identifies the primary technical hurdle as incorporating the full capabilities of large language models into the system while preserving efficiency. “We have two competing objectives,” he said. “On one hand, we aim to offer complete natural language flexibility, allowing users to ask anything. On the other hand, we need to maintain efficiency and use minimal resources, since processing thousands of feeds is incredibly demanding. Resolving this contradiction is the major technical challenge in our field, and it’s what we are intensely focused on solving.”