Background

A large part of the daily work at BM Systems involves receiving and handling various types of issue reports from customers. There are a few different categories of technical support that these reports demand, these can be handling of bugs, design changes and even user errors. The company also has a policy that response times should be under two hours.

Despite BM’s commitment to providing efficient and prompt support, there are several challenges. The biggest issue is the volume of incoming issue reports, making it difficult for support technicians to respond within the two-hour window while still providing meaningful assistance. Additionally, many of the issue reports are repetitive, often appearing in different variations, and a significant number of cases come with incomplete information from the customer.

One potential solution to these challenges would be to utilize AI, or more specifically a Large Language Model (LLM), this model could act as the first line of support, reducing the time burden on human support staff and ensuring that more complete information is available when a human takes over. This would create value both for individual customers, who would receive faster and more effective support, and BM as a whole.

Work Description

The work primarily involves investigating which type of model the student and BM wish to use as a foundation. Once the model is selected, the next step is to identify suitable types of training data, with previous support cases and manuals being strong candidates. It is also important to consider and research not only what data to use, but also how to apply it to the model—whether to use various types of data embeddings, direct training, or a combination of methods.

Following this, the focus should shift to developing the chosen model, primarily as a "Proof of Concept." At this stage, it is not necessary for the model to interact with existing issue report platforms. A crucial requirement for BM is that the model is developed and maintained locally, as it will be trained on sensitive data. To serve as a valid "Proof of Concept," it is also important to test the model in various ways, including running simulated cases with the support team at BM and possibly with one or multiple customers.

Once this stage is completed, the next step is to scale up the work that has been done so far and to consider how to integrate the model into existing workflows and platforms. This includes both the technical aspects of connecting it to the issue report system and how the interaction between support technicians and the AI should function, as well as how to keep the model updated and relevant over time.

If time allows additional aspects to consider include planning the integration with the existing workflow more thoroughly and potentially implementing a pipeline for continuous training. This would allow the model to evolve as more data is generated by the company, including additional support cases, manuals, and other documents that describe the systems.

The Student

We would prefer that you have experience with machine learning and AI, this experience might come either from relevant courses or personal interest demonstrated by home projects or similar. It is also preferable if the student is familiar with some of the following technologies, Python, Pytorch, TensorFlow, Keras or other relevant technologies and techniques.

Questions?

Contact Sebastian Sjöberg sebastians@bmsystem.se if you have questions regarding this Master’s Thesis.