AUTOMATING COMPLIANCE, AMPLIFYING SAFETY
WorkSafe: AI-Powered Mask Monitoring System
Background
—WorkSafe is an AI-powered compliance system developed with Veritek Engineering to help organizations ensure real-time mask adherence and workplace safety in the post-COVID era.
Our research is published in IEEE ICICCS 2022.
Role
UX Design · Front-end Development
Scope
8 Months · CAPSTONE
TEAM
1 UX Designer (Me), 3 Backend & AI Engineers
Figma
Balsamiq
Django
Overview
Client Brief
"Design a system that helps our organization enforce mask compliance post-COVID without constant manual oversight—making safety simpler and more consistent."
Despite government mask mandates, post-COVID workplace compliance stayed low—especially in close-contact settings. As a result:
We developed a real-time, AI-powered system that runs on each employee’s work device to monitor mask compliance without requiring manual supervision or external hardware.
Impact
The WorkSafe system, deployed at Veritek Engineering, significantly improved both safety and efficiency—even at a single pilot location.
%
increase in consistent mask usage
%
reduction in manual monitoring workload
%
Positive user feedback on ease of use
Workplace mask compliance was low, and manual monitoring wasn’t scalable. We built a real-time AI system that detects mask usage directly on employee devices, provides instant feedback, and sends violation logs to admins.
The result: better safety, less admin overhead, and high user satisfaction.
Research Process
Understanding the Broader Problem
To ground ourselves in the problem space, I started by reviewing national news coverage. Article after article pointed to the same patterns: inconsistent enforcement, public fatigue, and messaging breakdowns. Veritek’s issue wasn’t isolated—it reflected a nationwide challenge.
Understanding the PEOPLE BEHIND THE PROBLEM - USER Intreview
To get to the root of the issue, I designed a lightweight employee survey that combined Likert-scale and open-ended questions.
We distributed it across Veritek’s workforce (18–60 years old) and got responses from 42 employees—enough to spot behavioral patterns and breakdowns in policy perception.
I was specifically looking for three things:
Behavior: Are people actually wearing masks?
Belief: Do they think it still matters?
Breakdowns: Where’s the disconnect between policy and practice?
46%
Employess have contracted COVID-19 due to others not wearing masks
82%
of the time mask compliance check are not conducted.
71%
Employees do not wear their masks properly (covering nose & mouth)
Understanding the PEOPLE BEHIND THE PROBLEM - USER SURVEYS
To validate the patterns we saw in interviews and hear from more employees across roles, we launched a quick-turnaround survey of 42 staff members—including admins.
Our goal was to blend measurable trends with anonymous, honest feedback around sensitive behaviors like mask-wearing and enforcement.
What We Asked
Are coworkers consistently wearing masks correctly?
How often are mask checks conducted in your workplace?
How important do you believe mask-wearing still is?
Empathy map
To uncover what was driving low compliance, we mapped out both the admin and employee perspectives—what they were thinking, feeling, saying, and doing.
This helped us see the disconnects more clearly: admins felt overwhelmed, while employees felt disengaged.
Both wanted change—but for different reasons.
We used these insights to guide a more human-centered solution.
DEFINING OUR GOALS - WHAT WE SET OUT TO ACHIEVE
Talking to employees—through interviews and surveys—made one thing clear:
We didn’t need another rule.
We needed monitoring that was invisible, effortless, and respectful.
Our system needed to:
Automate mask detection in real time to reduce manual enforcement.
Empower users with visual feedback so they could self-correct (not be shamed).
Reduce admin burden and increase visibility—without increasing stress.
Only by aligning with what people actually felt—and why they behaved the way they did—could we design something that would actually stick.
WHAT SUCCESS WOULD LOOK LIKE
To track impact, we defined four key success metrics:
Compliance Rate — Are more people wearing masks correctly?
Admin Workload — Are fewer reminders needed?
User Engagement — Are users interacting meaningfully with the platform?
Feedback — Do people find the system intuitive and non-intrusive?
Conducted a full UX audit, 5 user walkthroughs, and affinity mapping to uncover major breakdowns in clarity, navigation, and messaging. Built personas and journey maps to reveal gaps in role-specific guidance—surfacing patterns that shaped the redesign priorities.
Defining the Problem
EXPLORING IDEAS: WHAT WOULD ACTUALLY WORK?
Before diving into features, we paused to ask:
What would actually help people stay safe—without wearing them down?
Quick surveys and chats made it clear: lapses came from fatigue, not defiance. Even the most diligent folks slipped up after hours on the job. They didn’t need rules or reminders—they needed less friction.
We explored a few early concepts:
Buddy systems for peer reminders
Interactive desk widgets to nudge behavior
Admin alert tools that pinged supervisors
They made sense on paper—but all relied on human intervention. And our users were already too exhausted for that
That was our turning point:
We needed a system that did the reminding for you, quietly and automatically.
THE SHIFT FROM ENFROCEMENT TO EMPOWERMENT
The new idea was simple—but powerful:
What if the system noticed unsafe behavior and helped users self-correct—without needing to call anyone out?
We knew our constraints:
No new hardware
No cloud processing
No breaking trust
So instead of external pressure, we focused on internal cues.
A real-time interface, just visible enough, that lets people course-correct themselves.
Design Principles We Locked In
This wasn’t surveillance. It was quiet reinforcement—nudging people toward community safety, not punishing them for slipping up.
USER JOURNEY MAP
Once we had our system concept, we zoomed out to map the full employee experience.
This journey map helped us spot key moments—where trust could build, where confusion might creep in, and how enforcement could feel fair instead of harsh.
It shaped every interaction to keep users engaged, not overwhelmed.
WIREFRAMING
With our principle clear—help users, don’t police them—we moved into wireframes.
We started with bare-bones flows to test: Would this feel respectful under real-world pressure?
EMPLOYEE UI
🟩 Green = OK
🟥 Red = Mask issue detected
Real-time webcam view with a color-coded frame.
ADMIN DASHBOARD
Shows mask-wearing trends by team or location
Never tracks individuals.
PENALTY & ESCALATION
After 3 warnings to fix their mask, users must pay a small fee to resume work.
Implementation
BUILDING THE SYSTEM: WHAT DID IT REALLY TAKE?
We knew what kind of experience we wanted — subtle, supportive, and secure.
But how do you actually build something like that?
Let’s walk through it.
Constraints
We had limited control over user hardware, no access to cloud hosting, and strict requirements around privacy and local data handling.
System Setup
We built the platform on Django for its clean structure, fast development, and seamless support for Python-based AI. It handled everything from user login to webcam access and real-time feedback.
Tech Stack
We tested YOLO, ResNet, and Transfer Learning, but chose MobileNetV2 + CNN for its speed, accuracy, and seamless Django integration
Accuracy
Our model reached 97.23% accuracy in detecting masks—more accurate than alternatives, and fast enough for live feedback.
The UX
Users log in, the webcam starts instantly, and they get clear green or red box feedback—aligning with our goal of helping them self-correct without feeling policed.
Implementation
CHALLENGES & OUTCOMES
We had to choose between CCTV and a mobile app—efficiency versus user-friendliness.
Real-time feedback also had to feel helpful, not harsh, which made UX decisions tricky.
Integrating the AI without slowing things down took work, and building user trust meant reframing the system as support, not surveillance.
This project pushed us to balance system complexity with real-world needs. We faced integration hurdles, UX friction, and user resistance—but solved them through iteration and empathy.
The Outcome?
+30% compliance:
Visual cues helped users self-correct.
–50% admin load:
Reports reduced manual monitoring.
Positive feedback:
Users found the system clear, respectful, and easy to use.
WHAT I TOOK AWAY