Conventional video analytics needs months of data collection, labeling, and training for each new use case. StreamVLM™ uses a Vision-Language Model that already understands images and language, so a detector you describe in a sentence starts working immediately.
The Long Tail
Every city has a list of things it needs to see that no analytics vendor sells a module for: a person lying on the ground, a damaged fence, illegal dumping behind a depot. Individually none of them justifies a training program. Together they are most of what a municipality actually worries about, and eight more are further down this page, each one written the way an operator would type it.
StreamVLM™ closes that gap. An operator writes the condition in plain English, the engine evaluates live camera frames against the prompt, and matches raise real-time alerts like any other event. Each camera channel supports multiple simultaneous prompt-based detectors, each with its own confidence threshold, alert cooldown, and event type.
StreamVLM™ is shipping in beta version on selected instances. It is included in the standard IREX price: there is no separate license, module, or subscription fee. The only additional cost is a local GPU node for VLM inference where a deployment requires one.
Example
"Alert when a person is lying on the ground." No dataset of people collapsing on platforms exists to train on, and collecting one would be neither practical nor decent. A Vision-Language Model does not need one: it already understands the scene, so the prompt is the specification.
Use Cases
Each card is a sentence an operator could type this afternoon, the row the platform writes back, and the person who still has to look at it. StreamVLM™ is shipping in beta on selected instances.
“Detect people with hands up or lying on the floor.” The platform reports the posture, not the crime. An event carries the frame, the camera and the time to whoever the branch nominates, and a person opens the clip and decides what to call it.
“Detect signs of a street robbery involving a two-rider motorcycle.” The event class says suspected on purpose. What the detector has is a pattern at a cash machine; the plate, the clip and the decision go to an officer.
See Two-Wheeler Analytics →
“Detect flooding in the underpass before the morning commute.” The same engine that watches people watches infrastructure. Water across both lanes raises an event before the first commuter reaches it, in time for a road team to close the approach.
“Detect graffiti being sprayed in the metro underpass.” A cost every transit operator carries and no vendor ships a module for. The alert is an event for a person to act on, and nothing in the detector identifies anyone.
“Detect snow and debris on the road if not cleared by 8 a.m.” A prompt can carry a deadline as well as a condition, so the detector stays quiet until the clearance window has passed. Here a ploughed ridge and a fallen branch still block a crosswalk and its ramp.
“Detect any camel on the highway.” Livestock on a carriageway is a regional problem with no off-the-shelf detector behind it. The sentence is the whole specification, and the alert reaches a road operator with the camera and the lane.
“Detect a person launching a drone.” Not a thing in the frame but something happening in it: open ground, a controller, and a quadcopter that has just left the ground. A person confirms it before anything else follows.
“Detect any usage of mobile devices, including phones and tablets.” In a corrections setting a handset is a security problem rather than a traffic one. The same prompt-defined detector applies, under the same Case ID and audit rules as every other module.
On the Map
One location at a time: the sentence an operator there would type, and the three rows the platform writes back for it. Some are prompt-defined StreamVLM™ detectors, the rest the trained modules and archive search they feed. Switch auto play off to read one, or open a match to see the frame at full resolution.
How It Differs
No data collection, no labeling, no per-use-case model. A new detector is a sentence, and it works from the moment you save it.
Designed to replace and unify several legacy modules, including fire detection, video-quality monitoring, and tamper detection.
Confidence threshold, alert cooldown, and event type are set individually, so a noisy condition does not flood the operator.
Every prompt, analyzed frame, alert, and agent action is logged with timestamp, operator identity, and Case ID, under role-based access.
StreamVLM events flow into the same three modes as classic modules: real-time alerts, investigations over the archive, and big-data export.
A prompt-defined detector sits inside the same ethics framework as the trained modules, including the narrow-constraints rule.
It is shipping in beta on selected instances. Whether a specific deployment is in scope is confirmed with IREX engineering rather than promised in advance.
No. StreamVLM is included in the standard IREX price with no separate license, module, or subscription fee. Where a deployment needs a local GPU node for VLM inference, that node is sized and priced with IREX engineering.
Several. A camera channel carries multiple prompt-defined detectors at once, each with its own confidence threshold, alert cooldown, and event type. IREX does not yet publish a per-camera ceiling: how the density scales depends on the customer’s hardware, and it is measured on your own cameras during the pilot.
It is designed to unify several of them, including fire detection, video-quality monitoring, and tamper detection. The trained high-frame-rate detectors remain the right tool for tracking people and vehicles at speed.
An open-weight vision-language model published by a third party and used as published: IREX does not fine-tune it or alter its weights, pins it by version and checksum, and serves it only on designated endpoints inside your instance, on the instance’s own GPU node. No frame, prompt or verdict leaves the instance, and no third-party cloud inference is used on any instance in any region. The model makes no tool calls, keeps no memory between frames, has no access to the archive, the event database, watchlists, other cameras or the network, and has no biometric function: it judges scenes and conditions and cannot identify a person. IREX does not publish which model it is; the identity, version and checksum are disclosed to a customer under NDA, because a named model is an attack surface and a fact that goes stale. Part B of the public AI Model Governance Policy sets all of this out.
Keep Reading
Bring the conditions your city actually worries about. If a prompt can describe it, we can usually show you a detector for it in the pilot.