AI-Powered Predictive Safety Workflows

Detail	Description
Role	Product Manager
Domain	Artificial Intelligence; Health, Safety & Environment (HSE)
The Big Picture	Transformed unstructured safety event data trapped in WhatsApp, emails, and legacy systems into predictive intelligence using AI-powered triage. The system analyzes observations and incidents to predict severity and identify leading indicators—enabling proactive hazard prevention. This strategic intervention increased safety event reporting by 61x, reduced processing time by 96% (45 mins → 2 mins), achieved 85% accuracy in high-severity risk prediction, and cut critical event response from 24 hours to 2 hours—moving the organization from reactive incident response to predictive safety management.
Key Skills Showcased	AI/ML Product Definition, Discovery, UI/UX Design, Stakeholder Alignment, Roadmap and Prioritization.

The Challenge: Latency in Proactive Safety

Construction sites generate vast amounts of safety data through observations (near-misses, hazards, unsafe conditions) and incidents (injuries, property damage, safety violations). However, this critical data was trapped in chaos—scattered across communication channels and legacy systems that couldn't transform it into proactive insights.

🚨 Problem 1: Unstructured Data Chaos

Safety event data—both observations and incidents—was scattered across WhatsApp groups, email threads, paper forms, and various communication channels. Critical safety information existed but was impossible to analyze, prioritize, or act upon systematically. HSE teams manually hunted through messages to find important events, with no structured way to connect observations to potential incidents.

🖥️ Problem 2: Legacy Systems Without Intelligence

Existing desktop safety recording systems were essentially digitized forms—no AI, no smart routing, no predictive capability. They captured data but provided zero intelligence. HSE teams manually reviewed every observation and incident to determine severity, urgency, and patterns. There was no way to identify leading indicators or predict which observations might escalate into serious incidents.

📊 Problem 3: Data Never Became Actionable Insights

Frontline workers felt their safety reports disappeared into a void. The feedback loop was broken—they reported hazards and near-misses but never saw outcomes or preventive action. Meanwhile, HSE teams spent 4-6 hours daily manually triaging events, with critical risks buried in hundreds of low-priority reports. Historical data existed but wasn't being used to predict and prevent future incidents. The fundamental issue: reactive safety management instead of predictive hazard identification.

Stakeholder Pain Points

👷 Frontline Workers: "We report hazards but nothing changes". Safety events lost in WhatsApp groups with no visibility into whether reports drive action.
⚠️ HSE Managers: Overwhelmed by volume, spending excessive time manually reading, classifying, prioritizing, and assessing data trapped in unstructured formats. This led to critical observations being missed or delayed and reactive responses instead of predictive prevention.
👨‍💼 Project Management Consultants (PMC): Lacked real-time visibility into leading risk indicators across the project, making it difficult to prioritize high-risk sites or allocate resources effectively.

The Solution: AI-Powered Risk Triage Engine

We built an AI-powered predictive safety intelligence layer that integrates with existing communication channels and systems to automatically structure, analyze, and triage safety events—transforming reactive data chaos into proactive hazard identification.

Workflow Comparison: Manual vs. AI-Leveraged

Workflow Step	Traditional Manual Process	AI-Leveraged Workflow
Data Capture	Reporter submits data via mobile form.	Reporter submits data via mobile app (text, photo, location).
Risk Triage	HSE Manager reviews the report, assigns Hazard, Severity, and Root Cause codes. (Time: 30+ min)	AI Engine performs real-time classification and prediction of Hazard Type, Severity (High/Medium/Low), and Probable Root Cause. (Time: < 5 sec)
Assignment	HSE Manager manually routes the observation.	System automatically assigns the observation to the relevant field supervisor based on AI-predicted severity and location.
Rectification	Reviewer rectifies the issue.	Reviewer acts on a pre-classified, high-priority task. HSE Manager only reviews high-severity, AI-flagged items.

Strategic Product Decisions

Decision 1: Intelligence Layer, Not Replacement System: Instead of forcing workers to adopt new tools, we integrated with communication channels they already used. The AI processes safety events from any source—preserving existing workflows while adding predictive intelligence. This "meet them where they are" approach drove the 61x increase in reporting.
Decision 2: Predictive Triage Over Comprehensive Classification: Started with intelligent severity prediction (High/Medium/Low) rather than attempting detailed hazard taxonomy. Simpler problem, faster value, higher accuracy. Once trust was established with 85% precision on critical events, more sophisticated classification could follow.
Decision 3: Human-in-the-Loop for Trust: AI handles high-confidence predictions automatically; uncertain cases route to human experts. This maintained trust during model training and allowed the system to learn from expert decisions, continuously improving prediction accuracy.
Decision 4: Visibility Drives Engagement: Built a closed-loop feedback system showing workers how their observations prevented incidents. This transformed reporting from "shouting into the void" into "being part of the safety solution"—driving sustained engagement and higher-quality reports.

My Role as Product Manager: Defining the Predictive Layer

As Product Manager at Navatech AI, I led the discovery, strategy, and product definition for this predictive safety solution. My focus was on deeply understanding operational realities and identifying the highest-value AI intervention points.

Discovery & Problem Validation: 🔍 Field Research & Stakeholder Engagement

Worked with ML engineers to define training data requirements and model performance targets
Shadowed safety teams to observe actual workflows and pain points first-hand
Analyzed existing safety data to identify patterns and gaps in current processes
Validated that the real problem wasn't reporting friction—it was the inability to extract predictive insights from existing data

Cross-Functional Collaboration: 🤝 Engineering & Operations Alignment

Conducted 15+ in-depth stakeholder interviews with HSE managers, site supervisors, and frontline workers across multiple construction sites
Collaborated with HSE domain experts to validate severity prediction logic
Partnered with operations and implementation teams to design closed-loop feedback mechanisms
Coordinated pilot rollout and gathered user feedback for iterative improvements

Key Insight that shaped the Product: "During discovery, I initially assumed workers needed easier reporting tools. But interviews revealed they were already reporting constantly—via WhatsApp, emails, verbal updates. The data existed. The problem was that this data was trapped in unstructured chaos and never transformed into actionable, predictive insights. This insight completely pivoted our product strategy from 'better data collection' to 'intelligent data transformation."

4. Outcome and Impact: Shift to Proactive Risk Management

The AI-powered predictive safety system transformed both the volume and quality of safety data while enabling proactive hazard identification—delivering measurable improvements across operational efficiency, incident prevention, and safety culture.

Key Measurable Outcomes

📊

61x

Increase in safety events reported—workers finally saw their data drive action

⚡

25%

Faster incident reporting and response time from event to resolution

✅

30%

Reduction in non-compliance events through early hazard identification

⏱️

50+

Hours saved per week for HSE teams—reallocated to proactive safety initiatives

⚙️ Efficiency Gains

Safety event processing: 45 mins → under 2 mins (96% reduction)
Reduced reporting time from 12 mins → under 3 mins (75% efficiency boost)
Manual triage time: 4-6 hours → 30 minutes daily
300+ daily events handled
Critical event response: 24 hours → 2 hours

🎯 Safety Performance

85% accuracy in high-severity risk prediction
Multiple potential incidents identified and prevented
Proactive hazard identification vs. reactive response
Data-driven safety decisions based on patterns

Cultural Impact: From Reactive to Predictive: Beyond the numbers, the system fundamentally shifted safety culture from reactive incident response to proactive hazard prevention:

Worker Engagement: Frontline workers became active participants in safety—they could see their observations preventing incidents, not disappearing into a void
HSE Team Empowerment: Safety managers shifted from firefighting to strategic prevention, using AI insights to identify and address systemic hazards
Operational Efficiency: AI reduced manual triage time by 96%, freeing HSE teams to focus on strategic initiatives
Data-Driven Decisions: Leadership gained visibility into leading indicators and could allocate resources proactively based on predictive insights
Continuous Improvement: The closed-loop system created a learning organization where every observation improved future predictions

Learnings & Next Steps

💡 Intelligence Unlocks Existing Data Value: The success of the Triage Engine validated the need for ML-driven workflows in operational environments. The primary learning was that user trust in AI is directly correlated with the accuracy of the high-severity predictions—it's better to have high precision (fewer false alarms) than high recall (catching everything).

💡 Discovery Prevents Building the Wrong Thing: Initial assumptions pointed toward easier reporting tools. Deep discovery revealed workers were already reporting constantly—the problem was data chaos and lack of predictive insights. Spending time understanding the real problem saved months of building the wrong solution. Always validate assumptions with users before committing to a solution.

💡 Trust Through Accuracy, Not Comprehensiveness User trust in AI correlates directly with prediction accuracy on critical decisions. High precision on important predictions (fewer false alarms) builds more trust than high recall (catching everything). We started with focused triage (High/Medium/Low) rather than comprehensive classification—simpler problem, faster value, higher accuracy. Scope small, nail it, then expand.

💡 Closed-Loop Feedback Drives Engagement Showing workers how their observations prevented incidents transformed engagement. When people see their input creates tangible outcomes, they contribute more and contribute better. Visibility into impact isn't a nice-to-have feature—it's fundamental to sustained adoption and data quality improvement.

💡 Predictive > Reactive in High-Stakes Domains In safety (and likely other high-stakes domains), the value isn't in better data collection or faster incident response—it's in prediction. The system's biggest impact was identifying leading indicators and preventing incidents before they occurred. Focus product strategy on where AI can fundamentally change outcomes, not just optimize existing processes.

Next Step: Extend the trained model to process video data (videos submitted by reporters) to automatically identify and classify hazards, further reducing reliance on human text input. Computer vision can detect unsafe conditions (missing PPE, scaffolding issues, equipment hazards) that reporters might not explicitly describe, further reducing friction and improving early hazard detection.

Interested in Learning More? Email or Connect on LinkedIn.