A San Francisco startup wants to turn every dusty security camera into something closer to an attentive employee. OpenVector, which emerged from Y Combinator's latest cohort, has built software that claims to transform passive surveillance into active automation using nothing more than plain English instructions.
The pitch sounds almost too simple: describe what you want monitored, and the system generates a custom AI model that watches camera feeds around the clock. No new hardware required. No complex programming. Just existing cameras suddenly able to trigger real business processes.
Andrey Gizdov and Vishal Urlam, the two co-founders behind the effort, are betting that computer vision adoption has stalled not because the technology isn't ready, but because deployment remains prohibitively complicated. Most systems on the market demand expensive proprietary cameras or require engineers to code elaborate rule-based logic. OpenVector sidesteps both obstacles by accepting natural language and working with whatever cameras are already mounted to the ceiling.
The company's homepage demonstrates the concept with scenarios that feel plucked from actual operations managers' wish lists: detecting bottlenecks at a manufacturing line and firing off WhatsApp alerts in 210 milliseconds, catching workers without proper safety gear, flagging unauthorized tailgating through secure doors—though these remain demonstration examples without publicly verified customer deployments. A car wash example shows the system recognizing a frequent customer, automatically applying a discount in the point-of-sale software, updating the CRM, and sending a confirmation text. All supposedly without human intervention.
Under the surface, OpenVector leans on academic research into something called foveated sampling. Gizdov, a Harvard PhD dropout and Fulbright Scholar, posted on LinkedIn in August that the platform processes 28 times less data per frame by sampling just 3 percent of pixels while maintaining equivalent output quality. The underlying work appeared as a Spotlight paper at CVPR 2025 in June, titled "Seeing More with Less: Human-like Representations in Vision Models." The implication: their vision model can monitor hundreds of camera streams simultaneously without choking on bandwidth.
Whether that performance holds up in messy real-world environments remains an open question. OpenVector has no publicly named customers or detailed case studies available as of September 2026.
Target Markets and Deployment
OpenVector is positioning the platform across manufacturing, logistics, warehousing, facilities management, restaurants, and automotive service. Use cases range from the mundane to the genuinely useful: ordering materials when staging areas run low, tracking item locations in sprawling warehouses, notifying restaurant hosts when tables open up, drafting service completion entries when vehicles roll out of repair bays.
The company is taking preorders for an offline processing station built on NVIDIA Jetson AGX Thor hardware. Pricing is currently listed at $449 monthly or $6,499 upfront, with a $100 refundable deposit holding a spot in line. Deliveries to California are slated for Fall 2026, though the company has stayed quiet on per-camera SaaS pricing for its cloud-based option.
Gizdov's co-founder, Vishal Urlam, studied embedded systems at Carnegie Mellon and previously worked on infrastructure at Tata. The team is small, just four people. A third-party LinkedIn post from UNI Network Group indicated the startup secured $500,000 from Y Combinator, consistent with the accelerator's standard investment, though OpenVector itself has not issued a funding announcement.
Crowded Field, Different Angle
The computer vision market is hardly empty. Ambient.ai pulled in $52 million in January 2022 for behavior-based security detection. Kibsi raised $9.3 million in June 2023 pitching no-code computer vision tools. Eagle Eye Networks, Spot AI, and Rhombus all offer platforms designed to squeeze intelligence from legacy camera infrastructure.
OpenVector's angle differs in emphasis. Instead of framing the product as security analytics, the company talks about vision-language-action workflows and direct integration into operational systems like CRMs, warehouse management software, and access control platforms. The language feels closer to robotic process automation than traditional surveillance software.
Prof. Shimon Ullman of MIT's EECS department offered a cautious endorsement on the company's about page: "OpenVector is exciting because it makes this technology work in real-time, for any camera, which can enable the automation of many types of labor."
That conditional phrasing captures the central uncertainty. Computer vision startups have promised effortless automation before. The question is whether OpenVector's natural language interface and lightweight processing truly eliminate the friction that has kept most companies from deploying these systems at scale, or whether the real barriers lie elsewhere. The company's reluctance to share customer names or detailed case studies makes it hard to judge traction beyond the demo videos.
For now, OpenVector is a bet that the gap between computer vision's technical capabilities and its practical deployment can be closed with better software design. Time will tell if that's enough.
