Twenty-one modules, each carrying an accuracy band we will commit to in contract.

Hardware-grade FRS terminals with depth and IR anti-spoofing at gates and high-traffic entries; software liveness for back-end matching.

• • Student attendance at gate
• • Staff and contractor entry
• • Visitor matching against pre-approved list

Accuracy

95% indoor adult under controlled lighting · 90% under-12 with periodic re-enrolment · 88% outdoor daytime

Hardware

Depth+IR FRS terminal at perimeter; NVR-attached cameras for back-end matching.

Caveat

Under-6 face recognition is not committed to as primary identity; we recommend RFID + parent app as fallback for that band.

Pre-registration, on-arrival capture, parental-consent workflow for visiting children, and badge issuance with auto-expiry.

• • Parent pickup with consent capture
• • Vendor and contractor flow
• • Patient attendant management

Accuracy

Process compliance >99% with mandatory-field forms; identity match per FRS module.

Hardware

Tablet kiosk at reception, ID-scanner attachment, on-prem print server for badges.

Read Indian commercial and private plates at vehicle gates — single-lane, multi-lane, and toll-style installations.

• • School bus arrival logging
• • Vendor truck entry
• • Employee parking allow-list

Accuracy

94% on standard high-security plates in daylight · 88% on aged/dirty plates · 85% at night with IR fill.

Hardware

Dedicated ANPR camera with shutter control; LED/IR illuminator for night; loop or radar trigger.

Caveat

Non-standard plates (handwritten, decorative scripts, partial obstructions) sit below SLA — we flag them for manual review rather than guess.

Two-factor entry for high-security zones: RFID badge presents, face confirms — one fails, both fail.

• • Server rooms and MDFs
• • Pharmacy and controlled-substance stores
• • Plant control rooms

Accuracy

98% true-accept with both factors present; explicit reject on any mismatch.

Hardware

Mifare/Desfire reader + depth+IR FRS terminal at door, with electric strike or maglock.

Two-stage detector: low-latency edge filter + cloud-of-one verifier, with mandatory human acknowledgement before escalation.

• • Firearm or knife visible on CCTV
• • Bag-content silhouette under known-good conditions

Accuracy

90% on visible firearms and large blades in well-lit CCTV; false-positive rate tuned to operator capacity, not vanity metrics.

Hardware

Standard 4MP+ IP cameras with sufficient pixel density on target zone; GPU at edge.

Caveat

Concealed weapons are out of scope. Walk-through magnetometers remain the right tool for that threat model.

Polygon-defined perimeter and restricted zones, schedule-aware, with separate models for human vs. animal vs. vehicle.

• • After-hours boundary breach
• • Restricted-zone entry during operations
• • Construction-site perimeter

Accuracy

93% detection on humans crossing defined polygons; <2 false alerts per camera per night with proper masking.

Hardware

Outdoor IP camera with IR; thermal recommended for monsoon and low-light reliability.

Zone-based dwell-time thresholds with operator-tunable durations; separates queue-waiting from anomalous loitering.

• • ATM lobby
• • School perimeter after-hours
• • Loading-dock area

Accuracy

88% on zones with consistent lighting and unobstructed lines of sight.

Hardware

Standard IP cameras; dwell-time logic runs on edge GPU.

Posture and motion-pattern analysis tuned to physical altercation, not animated conversation; always routed to human verification.

• • Hostel and dormitory common areas
• • Hospital A&E waiting
• • School playgrounds

Accuracy

85% on overt altercations within 8m of camera; lower in crowded scenes where bodies overlap.

Hardware

Wide-angle indoor IP camera with adequate frame rate (15fps+).

Edge-deployed acoustic models for gunshot, glass-break, scream, and distress sounds. Calibrated per-site to suppress local ambient noise.

• • Lab and workshop accidents
• • Storefront after-hours break-in
• • Hostel distress events

Accuracy

Class-dependent: 92% gunshot in semi-urban environments, 88% glass-break, 80% scream in noisy interiors.

Hardware

PoE-powered analytic microphone or audio-capable IP camera; on-prem inference, no cloud audio upload.

Configurable per industry — hard hat, safety glasses, masks, gloves, hi-vis vests, boots — with site-specific class definitions.

• • Shop-floor entry checkpoint
• • OT and clean-room compliance
• • Construction-site PPE audits

Accuracy

92% on standard PPE classes at entry checkpoints; lower in busy multi-class shop floors where we recommend zone-by-zone tuning.

Hardware

Choke-point overhead or eye-level camera with adequate pixel density on the target body region.

Visual smoke and flame detection as an early warning layer that complements — not replaces — fire-panel sensors.

• • Kitchen and pantry
• • Pantry-adjacent storage
• • Workshop and welding bays

Accuracy

85% on visible flame within camera FOV; smoke detection accuracy is highly scene-dependent and is treated as supporting, not primary.

Hardware

Standard IP cameras with no obstruction; integrates over BACnet/Modbus to fire panels where present.

Caveat

We never recommend replacing a code-compliant fire-alarm system with vision — only augmenting it for earlier visual confirmation.

Footfall and occupancy by zone, by hour, with per-site calibration. Counts are de-identified at the edge.

• • Capacity management
• • Footfall vs. conversion
• • Class and exam-hall headcount

Accuracy

97% at zone entries with overhead camera; 92% at oblique angles.

Hardware

Overhead 4MP camera at choke points; edge inference returns counts only, not images.

Aggregate dwell-time and traffic-pattern visualisation. All trajectories aggregated; no individual tracks persisted.

• • Store layout optimisation
• • Cafeteria and lounge usage
• • Campus circulation studies

Accuracy

Visualisation quality is qualitative — we publish methodology, not vanity scores.

Hardware

Overhead cameras across the analysed area; storage scaled to retention window.

Wait-time estimation, queue-length thresholds, and abandonment alerts — wired to staff escalation flows.

• • Billing-counter SLA
• • OPD waiting room
• • Cafeteria peak management

Accuracy

Wait-time estimates within ±20% under steady throughput.

Hardware

Standard IP camera with view of the full queue path.

Handoff of a tracked entity across overlapping camera fields, with behavioural threading for incident reconstruction.

• • Loss-prevention investigation
• • Plant-floor near-miss reconstruction
• • Campus incident review

Accuracy

Track integrity is scene-dependent. We publish per-deployment integrity numbers rather than a single headline metric.

Hardware

Existing camera grid with overlapping FOVs; edge GPU sized to camera count.

Forecast capacity peaks, staffing needs, and anomaly windows from historical occupancy and event data.

• • Roster planning
• • Peak-day staffing
• • Anomaly windows that warrant a human glance

Accuracy

Forecast accuracy is reported as MAPE per site after the first 90 days of training data.

Hardware

No additional hardware; runs against existing counting and occupancy data.

Density mapping for event safety and evacuation planning. Triggers tiered alerts before crush thresholds.

• • Assembly and prayer halls
• • Concourse and stairwell monitoring
• • Event venues

Accuracy

Density estimation within ±15% for medium-density crowds; higher uncertainty at extremes.

Hardware

Wide-angle overhead cameras; thermal supplements visual where lighting is poor.

Automatic recording with speaker tracking, board-content capture, and searchable transcripts. Indexed for revision and audit.

• • Lecture archive for absent students
• • Faculty review and CPD
• • Pedagogy audit

Accuracy

Transcript word-error rate around 8–12% on Indian-accented English; per-subject vocabulary tuning improves this.

Hardware

PTZ or fixed wide camera + ceiling boundary microphone; edge encoder.

Caveat

Recording requires explicit consent from faculty and a clear DPIA-backed policy on student visibility in frame.

Severity ranking, dedupe and suppression rules, escalation chains with on-call windows, and SLAs per alert class.

• • Reduce alert fatigue
• • Right-size P1 routing
• • Auditable response trails

Accuracy

We measure mean-time-to-acknowledge and false-positive rate per class, monthly.

Hardware

No additional hardware; runs on Ops Console infrastructure.

Aggregate behavioural baselining with privacy-preserving thresholds — flag deviations, never label individuals.

• • Unusual after-hours movement patterns
• • Shift-end anomaly detection
• • Site-wide drift indicators

Accuracy

Reported as deviation z-score against a per-site baseline; we do not claim absolute accuracy.

Hardware

Runs on aggregated counts and events; no new cameras required.

Caveat

We do not score individuals on this data. Anything that could constitute a discipline action requires human review and an appeal route.

Bespoke modules built to client requirements — laboratory protocols, equipment status, niche compliance checks.

• • Lab safety SOP audits
• • Specialised machine state detection
• • Regulator-mandated checks

Accuracy

Committed per module after a 4–6 week data study and pilot.

Hardware

Sized to the specific use case during Discovery.