Case Study
AI-Driven Monitoring and Decision Support for Experts at Razor Labs
Razor Labs is an Israeli deep-tech company specializing in AI-driven predictive maintenance for heavy-industry machines in sectors such as mining, energy, and manufacturing. Their technology integrates data from multiple sensors - vibration, temperature, pressure, current, and video - to detect early signs of malfunction, analyze causes, and prevent downtime.
The Inspector dashboard is an internal tool used by Razor Labs experts. It connects directly to the company’s data cloud and machine-learning algorithms, providing an interface for reviewing, validating, and improving AI-generated decisions. The goal: to transform complex multi-sensor data into clear, actionable insight for subject-matter experts and analysts.
The design challenge was to create a unified, intuitive interface that enables experts to:
To design such a system, we began by deeply understanding the people behind the process - the experts, analysts, and operators who interact with the Inspector daily.
Understanding the Experts Behind the System
Before defining any interface, the first step was to understand how Razor Labs' internal experts make decisions - what data they rely on, how they identify anomalies, and what prevents them from trusting the AI.
The research aimed to answer:
To uncover these insights, I used a mixed-methods approach combining qualitative and analytical research:
With data scientists, domain experts (SMEs), and customer success managers to map how decisions flow between teams.
Observing experts as they analyzed alerts inside internal prototypes.
Identifying where handovers or confusion occurred between algorithm, analyst, and manager.
Reviewing existing screens to pinpoint usability gaps and data overload.
These research activities revealed recurring patterns across different roles, helping define the personas, workflow gaps, and opportunities that shaped the product direction.
Understanding the People Behind the Decisions
Based on recurring behavioral patterns, responsibilities, and workflow similarities, I identified four primary user groups representing the core ecosystem around the Inspector platform.
Goal
Determine the root cause of machine anomalies as quickly and accurately as possible.
Frustrations
Needs
Goal
Translate technical findings into clear customer communication.
Frustrations
Needs
Goal
Review overridden AI decisions and continuously improve algorithm quality.
Frustrations
Needs
Goal
Understand which machines require immediate attention.
Frustrations
Needs
Across interviews, observations, and workflow analysis, several recurring patterns consistently appeared.
Experts switched between multiple systems to build context.
Excessive visualization made important insights difficult to identify.
Users hesitated to rely on predictions without understanding the reasoning behind them.
Analysts and SMEs documented feedback using inconsistent formats, making knowledge difficult to reuse.
These findings clarified that the challenge was not improving the AI itself, but improving how experts could understand, evaluate, and collaborate with AI decisions.
Defining the UX challenges that shaped the design vision.
The core idea behind the redesign was to create a single, intelligent workspace where experts could investigate anomalies, understand AI reasoning, and provide feedback - all in one place.
Instead of juggling between graphs, logs, and systems, users now navigate a streamlined interface that mirrors their mental workflow:
detect → validate → comment → improve
The design followed three key principles:
Simplify the experience
eliminate cognitive overload by focusing only on what's actionable.
Visualize reasoning
show why the AI thinks something is abnormal, with transparency and evidence.
Empower experts
design an interface that treats AI as a collaborator, not a replacement.
To translate research into structure, I mapped the user journey into three main stages.
Stage 1 - Detection
AI flags an anomaly, automatically prioritizing machines according to severity, confidence, and operational impact.
The interface immediately surfaces the most critical alerts, allowing experts to focus attention where it matters first.
Stage 2 - Investigation
Experts explore supporting evidence through sensor data, historical patterns, and AI reasoning.
Instead of navigating multiple disconnected systems, all relevant information appears within a single investigation workspace.
Stage 3 - Validation
Experts confirm, override, or dismiss AI decisions while documenting the reasoning behind every action.
This feedback continuously improves future AI performance while maintaining full transparency and traceability.
Machine Sensors
↓
AI Detection Engine
↓
Alert Prioritization
↓
Expert Investigation
↓
Decision & Feedback
↓
Continuous Model Learning
The redesigned workflow connects every stage of the decision-making process into one continuous experience.
Instead of forcing experts to jump between disconnected dashboards, information now flows naturally from machine signals to AI analysis, expert validation, and continuous model improvement.
Every interface decision was intentionally tied back to a research finding, ensuring that each design choice solved a real user problem rather than introducing additional complexity.
Hierarchical Information Structure
The interface prioritizes the most actionable insights first, while secondary information progressively unfolds as users investigate deeper.
Color Logic by Severity
A consistent severity scale helps experts instantly distinguish between low-risk observations and critical machine failures across every screen.
Contextual AI Explanations
Instead of exposing raw algorithmic output, the interface explains why anomalies were detected directly within the investigation workflow.
Persistent Investigation Context
Important findings remain visible while users navigate between sensors, historical data, and supporting evidence, reducing cognitive load and unnecessary context switching.
Lightweight Visual Language
A restrained visual system minimizes fatigue during long investigation sessions, allowing complex industrial data to remain readable over extended periods.
From Strategy to Interface
The following screens demonstrate how research insights were translated into a cohesive AI-powered investigation experience, helping experts quickly identify anomalies, understand AI reasoning, and make confident operational decisions.
The redesigned dashboard provides experts with an immediate overview of system health, active incidents, and operational priorities.
Critical alerts, AI confidence, severity levels, and machine status are organized into a single workspace, allowing users to understand where attention is needed before starting an investigation.
Once an alert is detected, experts transition from monitoring into investigation.
Instead of switching between multiple tools, the system consolidates AI reasoning, sensor evidence, historical context, and supporting insights into a single investigation workspace.
This allows experts to quickly understand the incident, validate AI recommendations, and investigate root causes without losing context.
After investigating the incident, experts review the AI recommendation before taking action.
Instead of treating AI as a black box, the interface keeps the human expert in control by supporting three possible outcomes - Confirm, Override, or Dismiss.
Every decision can be documented with supporting rationale, creating transparency, improving trust in the system, and continuously strengthening future AI recommendations through expert feedback.
Every expert action becomes part of a transparent decision history.
Confirmed, overridden, and dismissed incidents remain fully traceable, allowing teams to understand what changed, who made the decision, and why.
By capturing expert rationale over time, the system creates an auditable feedback loop that improves collaboration, builds organizational knowledge, and continuously strengthens future AI predictions.
Together, these workflows transform AI from a prediction engine into a collaborative decision-support system, enabling experts to investigate faster, make confident decisions, and continuously improve future model performance.