Spot AI's Iris: 20 photos and a conversation to build a camera agent
AI News reported on the April 2025 launch of a no-code builder that lets a plant manager describe a detection, feed it a handful of examples and attach an action such as shutting down a line. The claim was eight minutes, against the eight weeks a custom detector used to take.
Spot AI introduced Iris on April 9, 2025, during Google Cloud Next, and Muhammad Zulhusni's report for AI News the following day is one of the cleaner accounts of what the company was promising. Iris is an agent builder. A user describes, in conversation, what they want a camera to notice, supplies examples of correct and incorrect detections, and Iris trains an agent from that.
Eight weeks to eight minutes
The company's framing rested on a before-and-after. Rish Gupta, Spot AI's chief executive and co-founder, told AI News that specialized video detection used to require "dedicated AI/ML teams with advanced degrees, thousands of annotated images, and 8 weeks of complex development." Iris, he said, delivers "that same power in the hands of any business leader through simple conversation with 8 minutes and 20 training images."
Twenty images is a small training set, and the mechanism Zulhusni describes runs on feedback: the user marks detections as wanted or unwanted, and the agent adjusts. Spot AI's existing pre-built agents cover the common cases (a person in a restricted zone, a missing hard hat). Iris is for the cases specific to one plant or one store that no vendor would build a product around.
The examples in the article are the kind an operations manager would recognize. Product backing up on a line or fluid leaking under a machine in manufacturing. Boxes stacked in an unsafe way in a warehouse. Empty shelf space in retail. Telling staff from patients in a hospital. Verifying that safety gear is being worn. Once an agent exists it can be given an action to take, and the article lists shutting down equipment and locking doors among them.
What the report leaves open
Zulhusni's piece is a launch report and reads as one; there are no customers quoted and no independent test of the eight-minute claim. The engineering question of how a model trained on 20 images avoids firing on everything that looks vaguely similar goes unaddressed. So does the question of who is accountable when a self-built agent stops a machine.
What the launch did establish was the shape of the product category. By 2026, Brivo, Hakimo, Coram and Motorola were all selling some version of describe-it-and-the-camera-will-watch-for-it. Spot AI's version differs in insisting the user also teach it with examples and in attaching physical actions from the start, which I read as a bet that the buyer is a plant manager with a line to stop.
Read the original at AI News.