Development of an AI Chatbot — an Assistant for ATM Repair Experts

Even the best ATM repair expert cannot keep in mind all the situations that may occur when an ATM breaks down. That’s why company employees constantly turn to technical support specialists for help.

These specialists, in turn, help find the right information among numerous guidelines and navigate the various error codes that ATMs generate. However, every response from such a specialist takes time — and that means costs for the business.

That’s why the client decided to develop a chatbot to assist engineers: to reduce expenses, lower the workload on support, and speed up the search for answers.

The chatbot operates in Telegram, and all employee communication with it takes place inside the messenger.

An engineer can ask a question or simply send an ATM error code. In response, the bot provides a description of the error code, a technical repair guide, and, in some cases, a video.

We use OpenAI Embeddings to reduce response time and make queries more cost-efficient. This technology enables searching not across the entire database but only within the most relevant documents and their fragments.

The project is built according to the principles of “clean architecture,” which makes it easy to maintain the chatbot and implement changes to the product.

The client provided us with a massive dataset for further training of the neural network. The dataset included 200 MB of text information in DOCS, XLS, and PDF formats, as well as 30 GB of short video instructions.

To convert the videos into material suitable for LLM training, we transcribed them into text. We deployed Whisper AI, a transcription neural network, on our own server and transcribed all the videos.

Out-of-the-box LLMs don’t always work well with XLS format. Therefore, we converted all spreadsheets into JSON.

Next, we transformed all text into a vector format suitable for further training of the large language model. We used OpenAI Embeddings for this purpose.

The neural network had difficulty understanding user queries that included ATM error codes. To address this issue, we implemented code searches via SQL.

Additionally, the bot was reluctant to send users links to videos and documents, even when they were highly relevant to the query. To solve this problem, we created short descriptions for each file and video — after that, the bot reliably provided the necessary links.

Of course, we also automated this process — the descriptions for documents and videos were generated and added using the neural network.

Already, thousands of engineers are using our AI chatbot. As a result, the workload of support specialists has decreased fourfold. The bot helps resolve any typical issue, while operators only get involved in the most complex and unusual situations.

In the future, we plan to further train an open-source embedding model and deploy it on our own server. This will make the search for answers to user queries even faster and improve the quality of the results.