SightAgent™ · Continual interpretation

A VLM that knows what it is looking at.

SightAgent is Background's own trained Vision Language Model. Not a wrapper around a generic AI. A purpose-built interpretation layer trained to understand construction and infrastructure environments — people, plant, zones, proximity, activity and behaviour.

The distinction

Generic vision AI sees a person.
SightAgent understands the situation.

A generic vision model detects that a person is present. SightAgent understands that an unauthorised person is within a defined exclusion zone while a rotary piling rig is operating, the required supervisor is absent, and the control measures specified in RAMS-WP014-Rev C are unverified. That is not a detection. That is a situation.

The difference comes from training. SightAgent has been trained on construction and infrastructure environments — to recognise not just what is in the frame, but what it means in context.

Generic vision AI
○
Person detected
○
Vehicle detected
○
Movement detected
○
Object detected
Raw detections. No context. No meaning.
→
SightAgent™
●
Unauthorised person in exclusion zone
●
Rotary piling rig — operating state
●
Zone Z-003 boundary active — RAMS
●
Supervisor absent — CM-001 unverified
Typed evidence. Connected to context. Ready for CompoundOS™.

What SightAgent understands

Trained for construction.
Not adapted from something else.

SightAgent has been trained to understand the specific visual language of construction and infrastructure environments. The entities, activities, relationships and spatial configurations that define what is happening on a construction site.

01
People and roles
Workers, supervisors, visitors and unauthorised persons. Presence, absence, location relative to zones and plant. Whether PPE requirements are being met. Whether supervision is in place where the RAMS requires it.
02
Plant and equipment
Excavators, cranes, piling rigs, dumpers, rollers, MEWPs and more. Operating state — working, idle, moving, slewing. Position relative to people and zones. Whether plant is operating within defined boundaries.
03
Zones and boundaries
Exclusion zones, work areas, pedestrian routes, access points. Loaded from RAMS at shift start. Monitored continuously. Breaches identified and typed — whether the person is unauthorised, whether plant is operating, what controls should be in place.
04
Activity and behaviour
What is happening — excavation, lifting, piling, concreting, inspection. Whether the activity matches the method statement. Whether crews are productive, constrained or inactive. Cycle times, dwell, queue formation.
05
Proximity and interface
People-plant proximity. Distances relative to defined exclusion radii. Whether interface controls are in place. The difference between permitted proximity — a banksman with a working plant — and unsafe proximity without controls.
06
Time and pattern
Not just what is happening now — what has been happening over time. Dwell duration. Frequency of events. Pattern recognition across shifts, days and work packages. The difference between an isolated event and a systemic condition.

Edge and cloud

Two interpretation layers.
One intelligence picture.

SightAgent operates at the edge and in the cloud simultaneously. Each layer has a different job. Together they produce a complete interpretation of what is happening on site — immediately and over time.

Edge
Fast. Local. Real-time.

SightAgent Edge runs on the Background hardware unit installed on site. It processes camera feeds locally — without sending footage to the cloud — and produces typed observations in milliseconds.

Person detected — zone, location, state
Plant detected — type, operating state, position
Zone boundary — active, breached, clear
Activity — working, idle, constrained
Proximity — distance, interface type, controls
Processed locally. No cloud dependency for immediate interpretation. Network failure does not stop site intelligence.
feeds into
Cloud
Deeper. Contextual. Over time.

SightAgent Cloud reviews history, timelapse and documents to understand what happened over time. It connects edge observations to loaded context and builds the deeper picture that edge interpretation alone cannot produce.

Timelapse review — activity patterns over time
Document analysis — RAMS, permits, method statements
Trend detection — frequency, escalation, drift
Activity classification — what work package, what activity
Pattern recognition — across shifts, days, packages
Scheduled and triggered. Produces the deeper intelligence that feeds CompoundOS™ world models.

How SightAgent fits

Interpretation is the second step.
Context makes it meaningful.

SightAgent interprets what it observes. But interpretation alone is not enough. What makes an observation meaningful is the context it connects to — the RAMS, the programme, the zone boundary, the supervision requirement. That connection is what CompoundOS™ provides.

01
SiteRecorder™ captures
Camera feeds, documents, programme data — continuously captured and available to SightAgent.
→
02
SightAgent™ interprets
Raw camera feeds become typed, structured evidence. People, plant, zones, activity, proximity — understood and classified.
→
03
CompoundOS™ understands
Evidence combined with context. Capabilities evaluated. Understanding produced and delivered through IntelligenceInbox™.

SightAgent™

The difference between seeing and understanding is context. SightAgent provides the first. CompoundOS™ provides the second.

A camera sees a person near a machine. SightAgent understands that the person is unauthorised, the machine is operating, the zone is active and the supervisor is absent. That understanding is what makes intervention possible.

Next step

See SightAgent interpreting a real construction scenario.

Walk through how SightAgent interprets a site event — from raw camera feed to typed evidence to capability firing to intelligence delivered.

Book a demo How it works
Purpose trained — not adapted
SightAgent has been trained on construction and infrastructure environments. Not a generic model adapted for construction — a model built for it from the ground up.
Edge and cloud — both layers
Immediate interpretation at the edge. Deeper analysis in the cloud. Both layers feed CompoundOS™ with the evidence it needs to produce understanding.
Every observation is typed
SightAgent does not produce raw detections. Every observation is typed, structured and referenced — ready to be combined with context by CompoundOS™.