GenAI & Data Engineering Workshop [Part-1]: Enhanced Safety Assistant Through Intelligence

Reimagining Industrial Safety with GenAI and Data Engineering

Industrial safety depends on timely action and accurate information.

Our team set out to make both effortless through an AI-powered Industrial Safety Assistant that supports on-ground technicians with instant guidance, fault insights, and preventive measures, all powered by GenAI and Data Engineering.

This project combines RAG-based knowledge retrieval, contextual response generation, and real-time operational data – designed to work in fast, high-risk environments where every second matters.

The initiative was driven by a cross-functional team that brought together engineering depth and domain intuition:

→ Dipesh Bhavsar – Technical Project Manager

→ Vipul Makwana – Technical Leader

→ Palak Patel – Scrum Master

→ Ghanshyam Savaliya – Senior Software Engineer II

→ Alpesh Padariya – Senior Software Engineer II

→ Darshit Gandhi – Senior Software Engineer II

→ Swati Babariya – Senior Software Engineer I

→ Rency Ruparelia – Software Engineer

→ Bhavya Thakkar – Associate Software Engineer

→ Jil Patel – Associate Software Engineer

Together, they showcased how industrial safety can evolve from compliance to intelligence, where every query, alert, and recommendation is powered by data awareness and AI reasoning.

The Industrial Safety Crisis

The numbers that shaped this definition speak for themselves:

→ 2.8 million+ annual industrial injuries worldwide

→ 5190 daily workplace fatalities

→ $170 billion global economic impact

→ 60% of these accidents are preventable

When the team looked deeper, the manufacturing industry alone accounted for 403,000 injuries per year, construction reported 275,000, and mining showed an 87% higher fatality rate than average. The average cost per incident? $42,000.

For a domain where every second counts, delayed information or unclear instructions can directly cost human lives.

Root Causes Behind the Numbers

The team began by mapping the hidden factors behind these numbers. Their analysis uncovered four core causes:

1️⃣ Information Access: 45% of safety or operational guidance arrives too late.

2️⃣ Language Barriers: 67% of blue-collar workers face comprehension challenges with safety documentation.

3️⃣ Training Gaps: 38% report inadequate preparation for machine handling and risk response.

4️⃣ Emergency Response: 23% delays occur during crisis handling due to a lack of contextual direction.

This diagnosis helped the team define what their RAG system must solve – the gap between critical information and the worker who needs it, right when they need it.

The Opportunity of AI in Industrial Safety

Industrial safety may look like a compliance function, but the data shows a far larger opportunity.

The industrial safety market stands at $45 billion, growing at a 12% CAGR (2024–2030), with 78% of processes still manual.

In addition, the ROI potential of augmenting human capability with intelligent systems could reach 340% within just two years.

For enterprises, this is a chance to turn safety from a cost center into a measurable efficiency and productivity driver.

With the problem and potential clear, it’s time to step inside the build – the architecture, the logic, and the choices behind it.

Technical Specifications of GenAI-Powered Industrial Safety Assistant

The team engineered every layer to achieve real-time responsiveness and contextual accuracy in high-risk environments.

System Flow: From Manuals to Real-Time Machine Intelligence

Every industrial machine comes with a detailed manual, often buried in PDFs that are hard to navigate when a worker is already hands-on with the equipment.

The team addressed this friction by designing a seamless flow that converts static documentation into dynamic, conversational intelligence.

Here’s how the system works:

1️⃣  Document Ingestion: Admin uploads the machine manual (PDF). Text and images are extracted automatically.

2️⃣ Data Processing: Extracted content goes through chunking and detailing to preserve context. Each chunk is converted into embeddings and stored in VectraDB.

3️⃣ User Query Embedding: The worker’s query (e.g., “What voltage should I use for this AC?”) is also converted into an embedding vector.

4️⃣ Similarity Search: The system searches top-k matching chunks from VectraDB using cosine similarity for contextual precision.

5️⃣ LLM Response Generation: Matched chunks and user query are combined in a prompt and processed through GPT-4o, generating an accurate, guided response.

6️⃣ Recursive Function Calling: For incomplete replies (like “refer to appendix”), the system uses recursive function calls until a full, coherent answer is ready, including both text and relevant images.

Word Embeddings: The Language Behind AI

Embeddings turn words into numbers so machines can understand meaning, context, and relationships. Words with similar meanings appear closer together in this vector space.

Here, “man” and “woman” are close to each other, and so are “king” and “queen.”

Now imagine this: king – man + woman = queen.

This simple equation shows how vectors capture meaning – king minus man gives the idea of royalty, and adding a woman brings female royalty, leading to queen.

In our system, embeddings work the same way – they help connect machine parts, safety actions, and user queries with the right answers, based on meaning rather than just keywords.

From Query to Vector: Retrieving Semantic Representations

Every time a user asks a question, the system turns that query into a vector using the same embedding model applied to the documents.

The next step is finding which document vectors lie closest to this query vector, meaning which pieces of information share the same context.