Skip to main content

RESPOND · ALERTS & THREAT DETECTION

Know what matters, the second it matters.

Sentinel watches every camera and surfaces only what needs a human — a watchlist subject is on the operator’s screen in seconds. Every decision is reasoned, timed, and audited.

Command Center with a watchlist-match alert banner stacked at top showing camera, time and match confidence, and the threat dial at ELEVATEDClick to enlarge

Sub-5-second watchlist alerts · A live threat picture for the whole room · Every override audited.

Seconds
Watchlist subject to alert
from entering frame to the operator’s screen
5 levels
Auto-computed threat picture
LOW · GUARDED · ELEVATED · HIGH · CRITICAL
3 tiers
Per detection type
auto-alert, send to review, or dismiss as noise
Every override
Reasoned and audited
time-bound with a tamper-evident record
Watchlist alerts

The right face, the right plate, in seconds.

Add a subject to a watchlist and Sentinel surfaces the alert — with camera, location, time, and match confidence — fast enough to act on while it’s still happening.

  • Face and plate — alert on a person of interest or a vehicle/plate, across every camera at once
  • Sub-5-second design-target path from detection to the operator’s screen
  • Match confidence shown — the operator sees how sure the match is, never a bare yes/no
  • Stacked alert queue — top alert in full; the rest collapse into "+N more"
  • Act in place — View Event jumps to the clip; Track Person drops the subject on the Command Center map
  • Live monitoring chip — a constant, honest signal that the watch is active
Stacked alert banner: top watchlist hit with camera name, relative time, match confidence, and "+3 more alerts" expander plus Track Person and View Event CTAsClick to enlarge

The hit finds the operator — not the other way around.

Threat-level dashboard

One dial tells the room how tense the last six hours have been.

Thousands of detections distilled into a single, instantly readable indicator so a glance answers the only question a supervisor really has: is now a normal shift, or not?

  • Five levels — LOW through CRITICAL, color-coded, persistent across every page
  • Weighted scoring — watchlist hits, weapon detections, loitering, crowd, and anomalies each carry their own weight
  • Time decay — older events fade out of the rolling window so the dial doesn’t stay hot on stale activity
  • Trend at a glance — a rising / steady / falling arrow plus a six-hour sparkline beside the score
  • Recalculates on ingest — the dial updates as each event lands, no manual refresh needed
Threat-level banner at ELEVATED with rising trend arrow, six-hour sparkline, and contributing factorsClick to enlarge

Five levels, auto-computed, recalculated on every event.

Confidence thresholds

Decide what fires, what’s reviewed, and what’s just noise.

A three-tier confidence policy per detection type means the room sees only matches it can trust — high-confidence alerts, mid-confidence to review, the rest dismissed.

  • Three tiers per detection type — auto-alert (high confidence) · send to review (mid) · dismiss (below threshold)
  • Per-type control — face, plate, and other detection types each get their own policy
  • Confidence always visible — events carry their match confidence so operators prioritize the strongest matches first
  • Honest by design — a platform-wide confidence floor backs the face-pattern math; raising or lowering it is a deliberate, audited change
Three-tier confidence policy per detection type: face and plate rows each with auto-alert / review / dismiss bands and a confidence value on each eventClick to enlarge

Tune what fires per detection type — without touching a global dial that hides data.

Manual override

Human judgment outranks the algorithm — and leaves a record.

Command staff can pin the dial to a level by hand — but never silently: every override carries a required reason, an expiry, and a full audit record.

  • Pin the level — raise or lower the threat picture when command judgment supersedes the score
  • Reason required — no override without a written justification
  • Time-bound — set a duration (minutes to hours); the dial auto-reverts to the computed level on expiry
  • Fully audited — who overrode, to what level, why, and for how long — all recorded and replayable
Manual override dialog with level selector, required reason field, duration selector, expiry countdown, and an audit-log line formingClick to enlarge

Every override reasoned, time-bound, and written to the audit trail.

More detection capabilities

The full alerting picture.

Real-time alert banner

High-priority events ride the top of every operations screen — a stacked priority queue with the most urgent shown first and "+N more" behind it. Sound notification on a new alert (toggleable per operator). View Event and Track Person are built into the banner itself.

Instant escalation

Most of the time the threat picture should rise gradually. Some situations shouldn’t. Escalation rules let high-severity conditions push the posture immediately — bypassing the rolling score so a confirmed weapon alert or a cluster of watchlist hits lands as CRITICAL the instant it’s confirmed.

Counter-surveillance detection

A deterministic detector for subjects whose movement matches a reconnaissance pattern — a doubleback (A → B → A) or repeated loitering. Detections collect in a quiet inbox with severity sort, Acknowledge / Dismiss / Trace / Mute lifecycle, and full trigger-event citations.

Beta

Weapon detection

Weapon alerting is Beta — human-in-the-loop, multi-frame confirmed, never autonomous. A detection must persist across several consecutive frames before an alert fires. An operator must confirm or dismiss every weapon alert. No autonomous escalation.

How it works

From a detection to a decision.

1

Detect

An on-site AI agent processes each camera locally — faces, plates, people, vehicles, behavior, and the weapon signal (Beta) — each detection carrying a type, a confidence, and a timestamp.

2

Filter

Confidence thresholds decide what’s worth a human: high-confidence matches alert, mid-confidence ones go to review, the rest are dismissed — so the room isn’t drowned in noise.

3

Surface

A watchlist match raises the banner in seconds and sounds the room; the threat dial reweighs the last six hours and updates on its own. Serious conditions escalate the posture immediately.

4

Decide and account

The operator confirms, traces, or dismisses. Any manual override of the posture carries a reason, an expiry, and an audit record. Nothing acts on its own.

Under the hood

Specifications

Watchlist alert typesFace (person of interest) and vehicle/plate, across all cameras in the deployment
Alert latencySub-5-second design target from detection to operator screen (sized per deployment on reference hardware)
Threat levels5 — LOW · GUARDED · ELEVATED · HIGH · CRITICAL, color-coded
Threat scoringWeighted across the last 6 hours (watchlist hits, weapons, loitering, crowd, anomalies), with time decay; recalculates on event ingest
Threat indicatorsRising / steady / falling trend arrow + 6-hour sparkline; persistent banner on every operations page and the Command Wall
EscalationCondition-based rules that immediately set the posture, ahead of the rolling score
Manual overrideReason required, configurable duration (minutes to hours), auto-reverts to computed level on expiry, fully audited
Confidence policy3-tier per detection type (auto-alert / review / dismiss); platform-wide confidence floor MATCH_CONFIDENCE_THRESHOLD (default 0.6); deliberate, audited config change
Alert bannerStacked priority queue, "+N more" expander, toggleable sound (respects quiet / reduced-motion), per-alert dismiss, live monitoring chip
Weapon alert (Beta)Multi-frame confirmation (default 5 frames), higher confidence threshold (default 0.65), human confirmation required, no autonomous action
Counter-surveillanceDeterministic doubleback (A → B → A within 30 min) + repeated-loitering rules; trigger-event citations; ack / dismiss / trace / mute lifecycle; 90-day per-subject history
AccountabilityOperator actions and overrides audited; operator session capture & replay for after-action review
Scale1–5,000 cameras per deployment; alert, threat, and counter-surveillance queries scale-tested at 1.2M+ events with covering indexes

Underlying detection technology is detailed on the AI Vision Pipeline page. Performance figures reflect design targets on reference hardware; we’ll size your deployment with you.

See an alert turn into a decision.

Request demo access and we’ll send you a private, pre-loaded environment — take a live watchlist hit, watch the threat dial climb on its own, tune what fires, override the posture with a reason, and see counter-surveillance detection light up on a moving subject.