IREX Unveils StreamVLM™ — Video Analytics Defined by a Prompt

MURRIETA, Calif. — September 9, 2026 — IREX (IrexAI Inc.), the ethical AI company from California, today unveiled StreamVLM™, a Vision-Language Model (VLM) detection engine — shipping in beta on selected instances — that turns a plain-language description into a working video analytics detector.

For the past decade, adding a new capability to a video analytics system has meant commissioning a new AI model: collecting a dataset, labeling it, training, validating, and deploying — a cycle measured in months and budgets. StreamVLM removes that cycle. An operator describes a condition in ordinary English — “detect a person lying on the ground,” “alert when graffiti appears on a wall,” “identify flooding in the underpass” — and the platform begins watching for it on the selected cameras.

“Public safety agencies have never been short on things they need to see. They have been short on time and money to build a model for each one,” said Serge Smirnoff, Head of PR at IREX Inc. “StreamVLM changes who gets to decide what a camera network watches for. It is no longer a data science project. It is a sentence typed by the person who actually knows the neighborhood, the station, or the campus — and because it is IREX, every one of those sentences is logged, attributed, and reviewable.”

How it works

A StreamVLM detector is a named set of prompts with its own settings, applied to selected camera channels. Each prompt is independent, with its own confidence threshold, alert cooldown, and event type. Selected frames from live camera feeds are evaluated continuously against the prompts by a Vision-Language Model that understands images and language together, and matches generate real-time alerts within seconds.

Alerts arrive with a camera snapshot, a bounding box on the object of interest, timestamp, camera location, confidence score, and any extracted metadata, along with Play Video and Share actions. Every StreamVLM event enters the same pipeline as any other IREX analytics module, so existing workflows, dashboards, and integrations apply unchanged.

A single camera channel supports multiple prompt-defined detectors at once. A station camera can simultaneously watch for a person on the tracks, platform overcrowding, an unattended bag, smoke, fresh graffiti, and flooding at platform level. Detectors can be added or adjusted at any time without taking the system offline.

Beyond the pre-built module

StreamVLM sits alongside IREX’s specialized analytics modules for faces, vehicles and traffic, weapons, perimeter, rail and transit, crowds, fire, and camera integrity — and extends them into territory no fixed module catalog covers: infrastructure damage, illegal dumping, snow and ice hazards, unattended objects, unauthorized vehicles, worksite safety violations, non-standard signage and vehicle markings, and conditions particular to a single city.

Detectors created in StreamVLM are also available to Video Search & Investigations, IREX’s agentic investigation capability, which applies them across historical footage from thousands of cameras and compiles results against a verified Case ID.

Governed like any other detector

A detector that anyone can define in a sentence requires the same oversight as one that took a year to train, and StreamVLM is governed by the same mechanisms as the rest of the platform:

• Prompts are recorded. Creating, editing, and activating a detector are logged actions attributed to the operator who performed them, so what the system was told to look for is always reconstructable.
• Events are auditable. Every alert carries its channel, timestamp, confidence, and the prompt that produced it.
• Access is role-based. Granular permissions control who can create prompts, reach cameras, launch investigations, or export data.
• Investigations require a Case ID. Agentic steps are tied to an authorized case, creating a defensible chain of evidence.