CBRE SmartFM — AI-Driven Predictive Maintenance | Humayun Rashid
CBRE · SmartFM | 2022 – 2025 | Principal UX Designer

AI-Driven Predictive Maintenance
& HVAC Optimization

Transforming facility management from reactive firefighting to proactive intelligence

Traditional facilities management drowns engineers in alarm noise. At CBRE, I led the UX strategy to operationalize Ellis AI — a proprietary engine that reads thousands of IoT sensors, normalizes building system data, and surfaces only what matters. The result: a platform that reduced 70,000 monthly alarms to 1,400 actionable events, projecting $1.8M in cost avoidance across early-adopter clients.

98%
Alarm Noise Eliminated
$1.8M
Cost Avoidance Delivered
10%
PM Cost Reduction Projected
500+
Buildings Monitored
AI / ML Product Design Human-in-the-Loop IoT Data Visualization Condition-Based Maintenance Emerald Design System Figma Make Enterprise SaaS
The Problem

Turning Alarm Noise Into Actionable Intelligence

Traditional facility management is driven by intuition and reactive responses. Building engineers are overwhelmed by thousands of redundant, low-priority alarms generated by BAS systems that have never been properly filtered. Critical HVAC failures go undetected until something breaks — costing time, money, and occupant comfort.

Incoming BAS Alarms 70,000 / month
Ellis AI deduplication — 98% eliminated
After Deduplication 10,400 / month
Priority filtering — 10% deprioritized
Actionable Triage Events 1,400 / month
75% generate verified Work Orders · 25% closed remotely
Work Orders Created 1,050 / month
🔊
Alarm Overload
Engineers received 70,000+ events monthly — most duplicates or low-priority noise from BAS systems unable to self-filter.
🔧
Reactive-Only Maintenance
Maintenance ran on fixed schedules regardless of asset condition — missing failures in progress and wasting effort on healthy assets.
📊
No Single Source of Truth
Data lived across disconnected BAS, CMMS, and IoT platforms. No unified view meant portfolio-wide patterns were invisible.
My Role

Principal UX Designer — End-to-End Ownership

I owned the full design lifecycle for SmartFM's AI-powered maintenance intelligence features — from stakeholder research through shipped production designs, embedded with product managers, Ellis AI engineers, and facility management domain experts.

UX Strategy & Vision — defined the "single pane of glass" architecture for multi-application portfolio intelligence
AI Feature Design — designed CBM Insights, Ellis AI explanation panels, automated issue triage, and human-in-the-loop override flows
User Research — discovery with Facility Managers, ROC Engineers, and Account Leads to validate AI-driven feature assumptions
Design System Governance — contributed scalable AI interaction components to the Emerald Design System across web and mobile
Agile Collaboration — embedded in cross-functional squads (Earth, Water, Air, SoftServ), iterating sprint-by-sprint with engineering
Stakeholder Alignment — presented design direction to senior leadership, facilitated PI Planning, authored UX capacity frameworks
The Solution

Four AI-Enabled Design Breakthroughs

Operationalizing Ellis AI required translating opaque machine-learning outputs into experiences that facility engineers could trust, act on, and override. Each feature below represents a distinct design challenge in human-AI collaboration.

Feature 01 · Condition-Based Maintenance

Predictive Maintenance Logic

Instead of fixed maintenance schedules, I designed the UX for condition-based maintenance — where Ellis AI analyzes real-time IoT sensor data and historical issue patterns to predict failures before they happen.

The SmartFM Insights dashboard surfaces assets ranked by health score, estimated cost impact, and cost trend — giving engineers a prioritized action queue. The AI insight card provides root-cause context: not just "AHU 101 has issues" but "Insufficient airflow in 80% of downstream VAVs — two VAV units share problem code XYZ."

Key design decision: cost impact made visible at asset level. Showing $98,897 alongside AHU 117 transforms a technical alert into a business decision.

Condition-Based Maintenance
Feature 02 · Contextual Intelligence

Smart HVAC Insight Cards

The core design challenge: how do you communicate why the AI flagged an asset, not just that it did? An alert without context creates hesitation; context creates action.

I designed the Insight Card system — proactive cards that surface specific, machine-generated explanations with Recommended Actions and a direct "Create Work Order" CTA, collapsing the distance between insight and response to a single click.

Smart HVAC Insight Cards
Feature 03 · Explainable AI

Ellis AI Explanation Panel

The biggest barrier to AI adoption in enterprise operations is trust. If engineers don't understand why the system flagged an asset, they ignore the alert.

I designed the Ellis AI Explanation Panel — a side panel showing the exact asset ID, the specific threshold exceeded, and a historical date-value trend that proves the system's "Warning" or "Critical" rating with real sensor data.

This transparency layer — showing the reasoning chain, not just the conclusion — was the critical trust-building decision that drove engineer adoption.

Ellis AI Explanation Panel
Feature 04 · Human-in-the-Loop Automation

Automated Issue Triage & Remote Resolution

I designed the automation handoff layer — the moment where the system acts autonomously vs. when it surfaces a recommendation for human review.

Issue lifecycle states: New → Reviewed → Dispatched → Auto-closed. "Auto-closed" represents issues resolved remotely through software adjustments without dispatch. The SmartFM Explanation for every auto-closed event maintains a complete, human-readable audit trail — making automation visible and trustworthy.

CBRE Smart FM AI generated results
Design Process

Think · Make · Check

I utilized an iterative cycle to ensure Ellis AI's logic aligned with the mental models of Facility Managers, ROC Engineers, and Account Leads — three user types with fundamentally different needs from the same platform.

Think
Discovery & Framing
Conducted user research with FMs, ROC Engineers, and Account Leads. Used Google NotebookLM to synthesize stakeholder transcripts into actionable design direction. Ran SOS sessions and PI Planning workshops.
Make
Rapid Prototyping
Used Figma Make (vibe coding) to rapidly prototype intermediate states — where AI automation requires human validation before proceeding. Moved from insight to testable prototype in days.
Check
Validation & Governance
All AI-driven features audited against the Emerald Design System. Usability tested with real facility managers. Measured against UEM score, adoption rate, and engineer time-to-action.
Collaborate
Agile Delivery
Embedded across four squads (Earth, Water, Air, SoftServ), iterating sprint-by-sprint. Authored UX JIRA stories with acceptance criteria using Co-Pilot prompt templates.
Govern
HITL Framework
Established the human-in-the-loop governance model — defining precisely when Ellis AI acts autonomously, when it surfaces a recommendation, and when control returns to a human engineer.
Scale
Design System Contribution
Contributed AI-specific patterns to the Emerald Design System — confidence signalling, explainability panels, and AI status states — reusable by every product team building on SmartFM.
Outcomes & Impact

Measurable Results Across the Portfolio

Every design decision was validated against business metrics, adoption data, and direct user feedback from Facility Managers and Remote Operations Centers globally.

98%
Alarm Noise Reduction
From 70,000 monthly events to 1,400 actionable triage items through Ellis AI deduplication and filtering
$1.8M
Cost Avoidance
Delivered for early-adopter clients through proactive identification of critical HVAC failures
10%
PM Cost Reduction
Projected reduction through condition-based scheduling replacing fixed-interval maintenance
500+
Buildings Monitored
Active SmartFM deployment across 48 clients — CBM Insights covering 469 buildings across 11 accounts
53 → 71
UEM Score Lift
User Experience Metric lifted from 53.1 to 71.6 through iterative research-led design across sprints
44 → 57%
Product Adoption
Platform adoption grew as users shifted from skepticism toward active daily reliance on AI-generated insights
Reflections

What This Project Taught Me About AI UX

Trust is the product
The AI was technically sophisticated from day one. What took the longest to design was making it trustworthy. Every explainability feature, every confidence signal, every historical trend chart existed to answer one user question: "How do I know this is right?" Until that question was answered, adoption stalled.
Automation without visibility creates anxiety
When the system started auto-closing issues without explanation, engineers felt the platform was hiding things. Making every automated action visible and auditable — even when no engineer action was required — was the design decision that turned resistance into trust.
The handoff moment is the hardest design problem in AI
Designing when to hand control back to a human — and how to do it with enough context to act immediately — is more complex than designing the AI interaction itself. The intermediate state required more iteration than any other feature on the platform.
Data density ≠ complexity
Facility managers need dense, information-rich interfaces — not simplified ones. The instinct to reduce cognitive load through simplification often removed the context they needed to make decisions. The right answer was better organization and progressive disclosure, not less data.