AI Assistant in HSE: How to Implement and Use in Practice

Artificial Intelligence as a New Tool for the HSE Specialist

The introduction of digital technologies and artificial intelligence (AI) in the field of health, safety, and environment (HSE) is becoming not just a trend, but a necessity for optimizing routine processes. Tatiana Paklinskaya, an HSE specialist at the Edelweiss construction company, shares her practical experience of integrating neural networks into daily work. In her presentation, she reveals the "inner workings" of creating AI assistants that help automate the development of instructions and personal protective equipment (PPE) issuance standards.

From a Kitchen Analogy to Understanding AI

To effectively use neural networks, it is important to understand the basic principles of how they work. The speaker offers a simple metaphor: AI is the idea of cooking without a human, machine learning is the cooking method (baking, boiling), deep learning is a super multicooker (neural network), and a large language model (LLM) is the recipe and ingredients. The finished dishes are the familiar ChatGPT, GigaChat, and DeepSeek. Understanding these basics helps to correctly formulate requests (prompts) and get the expected result.

Creating AI Assistants for Specific Tasks

The presentation details the process of creating specialized AI assistants. The main principle that Tatiana derived is: "different tasks = different assistants". Attempting to create a universal helper leads to confusion and errors (hallucinations) of the neural network.

• Assistant for developing HSE instructions: Trimmed versions of regulatory documents (for example, part of the Ministry of Labor Order No. 772n) and local company regulations are loaded into the knowledge base. This allows the AI to focus only on the necessary rules and produce structured text by sections, bypassing limits on the volume of generated text.
• Assistant for selecting PPE standards: For this helper to work correctly, the speaker reworked the cumbersome tables from Order No. 767n, reducing the number of columns from nine to four. Loading responsibilities, special assessment of working conditions (SAWC) results, and equipment used allows the assistant to accurately select standards, leaving the specialist to only check the result and make targeted edits.

This approach reduces the time to prepare a draft instruction from several days to half an hour, where the main time is spent on the final formatting of the document.

Limitations of Neural Networks and How to Work with Them

The speaker shows by example that AI is a machine that requires clear instructions. The main limitations users face are:

• Output text volume limit: The neural network may arbitrarily shorten long documents. The solution is to configure the prompt to output text in parts.
• Need to structure the knowledge base: Loading unstructured data arrays leads to errors. Information must be pre-processed and "cleaned" of unnecessary details.
• Dependence on prompt quality: The more accurately and detailed the task is described, the higher the quality of the result. A detailed prompt can take up several pages, but it is what guarantees the correct operation of the assistant.

The Creative Potential of AI: Creating Educational Materials

In addition to routine tasks, neural networks open up new possibilities for creating engaging content. Tatiana demonstrated an animated electrical safety video created using AI. ChatGPT helped write the script and generate images in a unified style (3D Pixar), and GigaChat was used to "animate" them. This clearly shows how the role of a specialist is changing: from a simple executor to a process architect managing a digital team.

What You Will Learn from This Webinar:

• How to properly compose a prompt (request) for a neural network to minimize errors and "hallucinations"?
• How to prepare and structure regulatory documents (e.g., Order No. 767n) for loading into the AI assistant's knowledge base?
• Why is it necessary to create separate AI helpers for different tasks (instructions, PPE standards)?
• How to bypass the neural network's limit on the volume of generated text when creating large documents?
• How to use neural networks to create educational animated videos on HSE?