Trigon purchased a students’ machine learning project to make accounting software smarter 29.07.2025

Author:

Undercliff Creative

The company Trigon OÜ has decided to integrate into its accounting software a solution developed by students of a machine learning course at the University of Tartu Institute of Computer Science.

Trigon is an Estonian company developing accounting software with the aim of making accounting more standardised and efficient. According to the company's representative, Jaanus Karlson, the idea of creating a machine learning solution arose several years ago, when they began collecting a comprehensive dataset. Last autumn, they decided to offer their project to the students of a machine learning course.

At first glance, presenting the project assignment sounded technically complex. "Accounting is quite a confusing field for someone who has not encountered it before, especially at the data level," Karlson said. "I tried to approach it in a way that would help the students understand how the data is generated and how it is interrelated."

Machine learning model learned to classify invoices and tax codes

The aim of the project was to develop a machine learning model capable of recommending the correct company account and VAT code to the accountant based on purchase invoice data. This helps speed up accountants' work and reduces the risk of errors. Trigon provided the students with an extensive and very high-quality dataset - 240,000 digitised purchase invoices with an accuracy rate of 97-98%. In addition, the students received information about the workflows, such as how an accountant typically processes an invoice and how the chart of accounts functions.

According to the company, the collaboration with the university and the students went smoothly. "I did not really grasp how much work they had done until I started seeing very impressive probability scores," said Karlson. "We are very grateful to the university for helping to properly formalise the whole process with copyrights and agreements."

During the project, the student team particularly surprised the company by enriching the existing data with publicly available information - namely, the company's principal field of activity. Karlson praised the students' ingenuity, as this significantly improved the model's success rate. The final model was able to identify the correct account among the top five suggestions with over 90% accuracy, and performed even better with VAT codes.

"At one of the last meetings, when the students presented the model's success as graphs, I realised that the project had truly been a success," recalled Jaanus Karlson.

The purchase was a natural step

At the start of the project, the company internally agreed that if the solution proved good, they would buy it from the students. "Once we saw that the integration worked and the result was high-quality, there was no need for further discussion - it was a logical decision. We do not have that kind of expertise ourselves, nor the time to start learning machine learning from scratch," said Karlson.

Five months have passed since the purchase, and the solution continues to work excellently. According to the company's plans, they will start integrating it into the software in the second half of 2025.

Each year, dozens of companies submit their projects to the course. According to one of the lecturers of the course, Associate Professor in Artificial Intelligence Dmytro Fishman, both students and companies benefit from this course format. For students, it is a great opportunity to work on real-world problems from potential future employers. For companies, it allows them to achieve several goals at once: "Firstly, they receive a proof-of-concept solution for a problem that is important but not urgent enough to reach the development team's roadmap; secondly, they present themselves as a research-driven, forward-thinking company; and thirdly, they spot future talent whom they can potentially recruit after graduation," Fishman explained.

Companies are welcome to submit their projects to the machine learning course by 1 September.