Architecture, sensor fusion design, digital twin specification, and performance validation across eight AI engines for real-time occupancy sensing, utilization analytics, spatial simulation, conference room intelligence, hybrid workplace optimization, dynamic restacking, portfolio benchmarking, and predictive space planning — transforming $1.5 trillion in global wasted space into strategic intelligence.
Horizon deploys a multi-sensor fusion architecture that combines four sensor modalities — optical (computer vision), millimeter-wave radar, Bluetooth Low Energy, and WiFi signal analysis — to achieve occupancy detection accuracy that no single sensor type can deliver alone. Optical sensors with edge AI processing achieve 95% accuracy for people counting and presence detection using anonymous computer vision: images never leave the device, are instantly converted to encrypted occupancy signals (zeros and ones), and are immediately deleted. No personally identifiable information is stored, transmitted, or retrievable. mmWave radar detects presence and motion in small spaces without any visual capture, providing privacy-first occupancy detection for sensitive areas. BLE beacons track device proximity for zone-level density estimation, while WiFi scanning provides building-wide aggregate occupancy from existing infrastructure without additional hardware. The fusion engine weights each sensor modality by confidence level and spatial coverage, producing a unified occupancy map that updates every 2 seconds across the entire building.
The optical sensors use ELS (Edge-processing, Low-resolution, Secure) technology that processes all computer vision inference directly on the device. Low-resolution images are captured, processed through a lightweight CNN for people counting and active/passive classification, and immediately converted to anonymous numerical signals. The raw images are never stored on the device, never transmitted to any server, and are deleted within milliseconds of processing. The encrypted occupancy signals (integer headcount, boolean active/passive per zone, timestamp) are the only data that leaves the sensor. This architecture complies with GDPR DPIA requirements, ISO/IEC 27001, and SOC 2 Type II certification standards. Employee identities are architecturally impossible to derive from the sensor output — even with full access to the data stream, an adversary would receive only anonymous zone-level occupancy counts with no possibility of individual identification.
The fusion engine combines four sensor modalities using a weighted confidence model that accounts for each sensor's spatial coverage, accuracy profile, and environmental conditions. Open floor areas use optical sensors (high accuracy, wide coverage) weighted at 0.6, supplemented by WiFi signal analysis (0.25) and BLE (0.15) for redundancy. Small enclosed rooms use mmWave radar (0.7 weight) supplemented by BLE (0.3), providing presence detection without any visual processing. Lobbies and high-traffic corridors use optical sensors with DeepSORT cross-frame tracking for flow counting, weighted at 0.8 with WiFi as validation (0.2). The fusion engine resolves conflicts between sensor modalities using a Bayesian consensus model — if optical detects 12 people in a zone but WiFi detects 15 devices, the engine applies the optical count (more accurate for active occupancy) while flagging the WiFi surplus as potential passive occupancy from unattended devices.
Engine 02 transforms raw occupancy signals into actionable space intelligence by distinguishing between active occupancy (a person working at a desk) and passive occupancy (a laptop or bag claiming space while the person is elsewhere). This distinction is critical: nearly one-third of all desk time is passive occupancy, meaning that traditional badge-swipe systems that report 60% utilization may be measuring only 40% actual human presence. The engine computes peak/trough identification (Tuesday-Wednesday consistently 2-3× higher than Friday), 14-day demand forecasting at 89% accuracy, and occupancy-responsive HVAC and cleaning automation that eliminates servicing of empty floors. Engine 03 maintains a continuously updating digital twin — a live 3D model synchronized with real-time sensor feeds, HVAC telemetry, lighting states, and booking systems. The twin enables scenario simulation: test closing a floor, consolidating departments, switching to hot-desking, or adding a new team before making any physical change. BIM/CAD integration imports existing architectural models (Revit, AutoCAD, IFC) and overlays live operational data. Multi-campus federation supports 50+ buildings per deployment.
The active/passive classification model uses a multi-signal approach: optical sensors classify desk occupancy into four states — vacant (no person, no objects), passive-artifact (objects present, no person), active-present (person detected, actively working), and active-collaborative (multiple people at a single workspace). The classification runs at the edge on the sensor device, adding negligible latency to the occupancy signal. Aggregate active/passive ratios are computed per zone, floor, and building, with trending analysis that reveals behavioral patterns: departments with high passive ratios may have employees who prefer to work elsewhere but leave belongings to "claim" desk space — a behavioral signal that hot-desking or unassigned seating could be introduced without resistance. The active/passive distinction transforms utilization metrics from misleading artifact-detection into genuine human presence measurement, enabling space decisions grounded in reality rather than inflated utilization numbers.
The digital twin implementation uses a three-layer architecture: (1) the geometric layer, imported from BIM/CAD (Revit, AutoCAD, IFC 4.0) and stored as a lightweight 3D mesh optimized for real-time rendering in web browsers; (2) the semantic layer, which maps every zone, room, desk, and amenity to its functional classification, capacity, departmental assignment, and equipment inventory; (3) the live data layer, which overlays real-time sensor feeds (occupancy, temperature, CO2, humidity, lighting) onto the geometric model with 2-second synchronization. The what-if scenario engine operates on a copy of the semantic and live data layers, enabling simultaneous testing of multiple scenarios (close Floor 7, move Marketing to Floor 3, introduce 1.5:1 hot-desking for Engineering) with projected impact on utilization distribution, energy consumption, employee travel paths, and amenity demand. Scenarios complete in under 3 minutes for buildings up to 500,000 square feet and produce side-by-side comparison dashboards showing current state versus proposed state across every measured dimension.
Engine 04 monitors actual conference room occupancy in real time, detects no-shows within 12 minutes of a booking start, and automatically releases the room back into available inventory. Industry data shows that 40–50% of booked conference rooms are either no-shows or end more than 15 minutes early, while employees spend an average of 12 minutes per meeting searching for available rooms. Right-sizing analysis reveals that 65% of meetings in 12-person rooms have 3 or fewer attendees — meaning enterprises are heating, cooling, and maintaining large conference rooms for small conversations that would be better served by huddle spaces. Engine 05 optimizes hybrid workplace design: hot-desking ratio calculations per department based on actual attendance patterns, neighborhood zoning that co-locates teams on their peak days while consolidating quiet floors on low-demand days, and occupancy-responsive building operations that reduce HVAC, cleaning, and lighting on floors with zero or minimal presence. As of late 2025, 70% of job seekers prioritized hybrid work flexibility, even though only 24% of job postings offered it.
The conference room intelligence engine operates on a three-stage detection pipeline: (1) pre-meeting validation — 15 minutes before a booking, the system checks whether the organizer's calendar shows an active conflict or travel status that makes attendance unlikely, sending a proactive confirmation request; (2) no-show detection — if zero occupancy is detected 12 minutes after the booking start time (via room-mounted mmWave radar or optical sensor), the room is automatically released and the organizer is notified; (3) early-end detection — if the room empties 15+ minutes before the booking end time, the remaining slot is released for walk-up or ad-hoc use. Over time, the system learns individual and organizational booking patterns: serial no-show bookers receive gentle behavioral nudges, while chronically overbooked rooms are flagged for policy review. A deployed enterprise reclaimed an average of 4.2 hours of daily conference room availability per floor — equivalent to adding 2 new conference rooms without construction.
The hybrid optimization engine computes optimal desk-to-employee ratios at the department level by analyzing actual attendance patterns over rolling 90-day windows. Engineering teams with 82% Tuesday-Thursday attendance and 34% Friday attendance can operate at a 1.2:1 desk ratio (120 desks for 100 engineers) — but Sales teams with 45% average attendance and no clear peak day can operate at a 2:1 ratio. The engine prevents the oversimplified "one ratio fits all" approach that fails in practice by computing per-team ratios, then optimizing neighborhood zoning to co-locate teams that share peak days on the same floors. On low-demand days (typically Monday and Friday), the system recommends consolidating all occupancy onto a subset of floors and powering down the rest — reducing HVAC, lighting, elevator service, and cleaning to only the floors with active presence. An enterprise deployment achieved an 18% increase in return-to-office attendance after optimization — not through mandates, but by making the office worth coming to on the days people were already there.
Engine 06 optimizes department-to-floor assignments using a constraint-satisfaction algorithm that maximizes inter-team adjacency (placing collaborating teams on the same floor or adjacent floors), respects physical constraints (secure areas, specialized equipment, ADA accessibility), and minimizes move disruption through phased migration planning. Engine 07 normalizes utilization data across the entire portfolio — adjusting for building age, market, industry, and employee density — to enable true apples-to-apples comparison across sites and against industry benchmarks. Engine 08 is the platform's most strategically valuable capability: predictive space planning that models future space requirements based on workforce growth forecasts, hybrid adoption trajectories, utilization trend lines, and planned organizational changes. The engine answers the question every CFO asks: "Do we actually need more space, or do we need to use what we have differently?" In validated deployments, Horizon has prevented over $180 million in unnecessary construction and lease commitments — including a university where a provost discovered that a planned $42 million science building was unnecessary because existing laboratory space was utilized at only 31% during non-peak hours.
For every proposed construction, expansion, or new lease, the predictive space planning engine calculates a Capital Avoidance Score — the probability that optimization of existing space can deliver equivalent capacity without new construction. The score integrates five dimensions: (1) current utilization headroom — how much unused capacity exists across the portfolio after accounting for active vs. passive occupancy; (2) hybrid absorption potential — how many additional employees can be accommodated by implementing or increasing hot-desking ratios based on demonstrated attendance patterns; (3) conference room recovery — how much capacity can be reclaimed through no-show elimination, right-sizing, and conversion of underused large rooms to bookable huddle spaces; (4) schedule optimization — how much capacity can be created by consolidating peak-day demand onto fewer floors and powering down the rest; (5) layout reconfiguration — how much capacity can be gained by converting private offices to shared spaces or redesigning open floor plans based on actual collaboration patterns. A score above 0.7 triggers a mandatory optimization review before any capital expenditure is approved.
The growth absorption model projects how many additional employees the existing portfolio can absorb through optimization before any new square footage is required. The model processes workforce planning data (HR growth forecasts by department, location, and timeline), hybrid policy parameters (mandatory in-office days, department-specific policies), historical utilization trends (trajectory analysis showing whether utilization is increasing, stable, or declining), and organizational change plans (mergers, divestitures, department launches, office relocations). The output is a capacity timeline: "At current growth rates and projected hybrid patterns, the Chicago office can absorb 340 additional employees within existing space through 1.5:1 hot-desking, floor consolidation, and conference room right-sizing. New space will not be required until Q3 2028." This intelligence transforms the CRE function from a reactive real estate broker into a strategic capacity planner — anticipating space needs years in advance and preventing the panic leasing that drives enterprises to sign expensive long-term commitments for space they may never fully use.