Active risk intelligence / v4.2

Most incidents
announce themselves
before they happen.

Your operation already emits thousands of risk signals an hour. Almost none of them are read together. Foresight is the correlation layer over the signals you already generate, surfacing the seconds that matter while someone can still act on them.

No new infrastructure On-premise inference One model across every signal
Detection library / one model family

One model family, a spectrum of events

Most risk tools are built for a single job: a people counter, a door alarm, a loss-prevention system. Foresight runs one shared model family across every signal a site produces, so the same deployment that catches an unstable transfer also catches a propped fire door and an after-hours badge anomaly.

Critical

Falls & gait instability

Pose estimation tracks gait symmetry, transfer mechanics, and reach-for-support behavior. Most flagged events fire before contact with the floor. The value is the near miss, not the incident report.

  • gait_asymmetry
  • unassisted_transfer
  • floor_contact
  • wall_support_seek
Critical

Aggression & assault

Escalation is a motion signature before it is an act: closing proximity, raised limb velocity, defensive posture, bystander recoil. Foresight scores the trajectory and alerts during the buildup.

  • proximity_breach
  • velocity_escalation
  • defensive_posture
  • crowd_recoil
Critical

Powered equipment & lift injuries

A tip-over is legible before the truck goes over. Steer angle against travel speed with the mast raised pushes the combined center of gravity outside the stability triangle, and Foresight scores that envelope continuously. Pedestrian intrusion into an equipment aisle is treated as the same class of event.

  • stability_envelope
  • travel_mast_raised
  • pedestrian_aisle_breach
  • lift_posture_strain
Elevated

Theft & shrink

Concealment gestures, dwell anomalies at high-value fixtures, and unauthorized removals, correlated across zones and systems so a single actor moving through four areas reads as one incident.

  • concealment_gesture
  • dwell_anomaly
  • asset_removal
  • multi_zone_track
Elevated

Environmental hazards

Spills, obstructed corridors, equipment abandoned in a walking path, cluttered egress routes. Standing conditions get a persistence timer, so a hazard unaddressed for 20 minutes escalates on its own.

  • surface_moisture
  • egress_obstruction
  • path_clutter
  • persistence_escalate
Monitored

Access & perimeter

Tailgating at controlled doors, propped exits, after-hours presence in restricted zones. Reconciled against your access-control log so a badge-in with three bodies is flagged as one.

  • tailgate_detect
  • door_propped
  • afterhours_presence
  • acs_reconcile
Architecture / edge-first

Raw signal never leaves the building

Inference runs on an appliance in your own rack. What crosses the network is a stream of scored, de-identified events: a class, a confidence, a zone, a timestamp. That single decision is what makes the system deployable in regulated environments.

SITE BOUNDARY raw signal stops here Existing signals Edge appliance Signal fusion Risk scoring Active alert vision · access · IoT on-site inference cross-system severity × context routed to responder raw feeds events only correlated scored responder outcomes tune thresholds per site
The boundary is the design. Everything left of it is yours and stays yours; only scored events cross.
Connect

Point at what you have

Standard protocols across vision, access control, and building systems mean most sites are streaming into the appliance the same afternoon. No rip-and-replace, no new cabling.

Calibrate

Map the zones

Two weeks of baselining learns your traffic patterns, shift rhythms, and normal dwell. Thresholds are set per zone, not per product.

Detect

Score in real time

Signals pass the model family continuously. Events are scored on severity, confidence, and context. A fall risk at 3am scores above the same signal at noon.

Route

Alert the right person

Escalation paths by zone, class, and shift. The alert carries a five-second clip and a location, so the responder already knows what they're walking into.

Signal strategy / use what's there

You already generate the signal

A mid-sized facility produces a continuous stream of operational telemetry: presence, motion, access events, door states, environmental readings, call and dispatch records. Each system watches its own slice and none of them talk. That is a sensor network with no analytics layer. Foresight is the analytics layer.

Baseline

The infrastructure you own

Foresight reads existing systems in place, on standard protocols. Nothing gets replaced, and equipment more than a decade old routinely carries production detections.

Fusion

Correlation is the product

One signal is ambiguous. A door event, a presence anomaly, and an environmental reading in the same zone within ninety seconds is a described situation. Foresight scores the combination, which is why it flags what any single system misses.

Where
needed

Strategic sensing partners

Some signals genuinely need physics your current stack doesn't have. We maintain co-engineering agreements with specialist manufacturers across five sensing categories, specified only where a site's risk profile earns the spend.

Never

Wearables on people

Pendants get removed, batteries die, and compliance is a daily fight. Ambient sensing has no adherence problem.

Signal sources ingested

Typical deployment reads four to seven of these. All of it is infrastructure most sites already run.

Vision
existing fixed and PTZ estate
Access control
badge, door state, propped alarms
Building systems
HVAC, water, temperature, power
Call & dispatch
call bells, panic, radio events
Presence telemetry
Wi-Fi / BLE density and dwell
Operational logs
work orders, incidents, staffing
Point of sale
voids, overrides, shrink patterns

Supplementary, where specified: thermal · mmWave radar (no image captured) · acoustic array · depth / ToF · edge accelerators

Governance posture

Written to survive a regulator's question, not just a security questionnaire.

Inference
on-premise, air-gappable
Egress
event metadata only
Identity
no face recognition, ever
Masking
per-zone, enforced at capture
Retention
configurable, 24h default
Audit
immutable access log
Privacy & governance / non-negotiable

Detection without identification

Foresight classifies behavior, not people. The model has no face recognition capability and no persistent identity graph. A tracked subject is a numbered track that expires when they leave the frame. This is an architectural choice, not a setting.

Zones

Masked at capture, not in post

Bathrooms, bedrooms, and any zone you designate are excluded before a frame reaches the model. There is no recording to leak because there is no recording.

Labor

Not a productivity tool

Foresight does not score staff, time tasks, or produce per-employee reporting. Deployments in union environments hold because that boundary is contractual, not aspirational.

Evidence

Discoverability is designed for

Short retention windows and event-level logging keep the record proportionate. Your counsel gets a defensible policy instead of an archive of everything.

Deployments / where risk concentrates

Wherever seconds change the outcome

Foresight is licensed directly by the organizations running these sites, as an operating system for their own risk management. What it costs them is measured in incidents avoided, staff response time, and conditions corrected before anyone is hurt.

Senior living

Skilled nursing, assisted living, CCRC

Falls, aggression, and standing hazard conditions drive the severity curve. Foresight's deepest detection library sits here.

Retail

Multi-site specialty & grocery

Organized retail crime patterns and slip-and-fall exposure in the same deployment, correlated across a district.

Industrial

Warehouse & distribution

Pedestrian-forklift conflict, blocked egress, PPE compliance zones, and dock-door access control.

Hospitality

Hotels & gaming

Guest-area incident detection with strict privacy masking, plus back-of-house safety and access monitoring.

Education

Campus & residential facilities

Perimeter integrity, after-hours presence, and aggression detection in common areas without identifying students.

Portfolio

Multi-site operators & risk pools

Aggregate, de-identified risk telemetry benchmarked across a portfolio, so a group can see which sites are drifting before an incident says so.

Get started / site survey

Bring us one building

We start with a survey of a single site: an inventory of the signals you already generate, zone mapping, and a read on which event classes actually apply to your operation. You get a deployment plan and an honest answer about what your existing infrastructure can carry. Four weeks, one facility.

Typical response within one business day.
Site surveys scheduled 2–3 weeks out.