Physical security is a universal requirement for all organizations, which in the U.S. alone deploy an estimated 85 million security cameras to monitor their premises against threats and crime.  Yet almost all the exabytes of video data generated by these cameras serve no purpose: They vastly exceed the attentive capacity of people to view them.  Indeed, experiments show that even dedicated personnel cannot effectively detect threats or illicit behavior while viewing multiple screens simultaneously over time.  Most surveillance footage is deleted unviewed.

This, among many other reasons, is why we are honored to lead the Series A of Coram AI, an AI-powered video security and monitoring platform founded by two accomplished technologists in video AI, Ashesh Jain and Peter Ondruska.

Video AI enables organizations to install an attentive watchdog behind every camera at dramatically lower cost and greater vigilance than the human equivalent. The same advances that have transformed machine understanding of text and images are being applied to video data, with similarly remarkable results: It is now possible to query video data sets in natural language and find matching clips in seconds that would have taken humans hours to identify. Prompts such as “Show me any time two people carried a box within ten feet of a forklift,” or “Trace the path that a person carrying a yellow backpack walked in this building yesterday,” are now intelligible prompts to video AI models.

Not only does this render existing surveillance cameras much more effective, it can also enhance privacy and coverage at the same time: AI agents can be trained to identify only threats and problematic behavior across many more endpoints.

That’s great in theory, but the execution turns out to be much harder.  Video processing at scale with dozens of cameras or more is computationally intense, compounding the potentially expensive computing demands of the AI models themselves.  False positives and false negatives are dangerous and can threaten privacy, so precision and low latency are essential.  There are dozens of security camera manufacturers, with incompatible signal types and technical specs, and many organizations have existing implementations of third-party cameras that they don’t want to fiddle with or reimplement.  These and other challenges confound the potential of video AI — which is where Coram comes in.

Ashesh and Peter recognized that the huge horizontal market for AI-powered video security would only be unlocked with cost-effective, consumer-grade solutions that worked with existing cameras and that continuously improved in the precision of their detections.  Trained in the demanding domain of self-driving vehicles at Lyft, where multi-model sensor data and computer vision have to be integrated in real-time, these two founders combined their academic understanding of machine learning and video processing with pragmatism about commodity hardware and cameras.  Within a short period of time since founding, they have come to market with a commercially compelling platform already trusted by many businesses and schools to protect their premises.

As large as this market is, the future holds considerable potential for expansion.  Industrial and ergonomic safety are closely adjacent to physical security.  Specialized detections can be connected to site access control systems, and the technology can send real-time alerts to law enforcement about theft or intrusion, or to site managers when safety hazards are spotted.  Video models can be expanded to incorporate other sensor data for leak, temperature, or smoke detection.  Operational metrics like space usage, customer traffic, and attendance can be added to the set of detections manageable through a single interface.

We believe these kinds of capabilities will evolve video monitoring from a mere deterrent or passive documenter into a real-time operational input.  We are proud to partner with a company on the vanguard of making our world safer and more efficient in a privacy-preserving way.