One software for all documents

Sometimes computers have blind spots, such as when they are asked to recognise documents that have not been structured specifically for them. “Computers read from top left to bottom right. They don’t recognise line spacing or boldface type,” says Tim Beck, CEO of ETH spin-off BLP Digital.

Tables are a big problem for computers, because the values in a table are only meaningful if their position in the table is known; that is, their respective row and column. Beck and his team set out to teach computers to recognise tables and any other semi-structured information in documents.

Invoices, delivery slips, expense reports, leases, CVs, emails: the range of applications for this type of software is huge. Beck sees enormous potential here. One major Swiss health insurer, for example, processes 13 million invoices annually. When a wholesaler orders bananas from South America, the process will generate 17 documents by the time the fruit reaches shops in Switzerland.

Software improves over time

“We want to automate more of these processes and make employees’ jobs easier,” says Beck. The 28-year-old’s long-term vision is to design a fully versatile piece of software that enables the flawless automation of all common document types through machine learning. The software is designed to improve over time as it collects greater volumes of data. This ambitious goal is, of course, still some way off, but Beck’s younger brother and co-founder Sven says it is important to keep this goal in mind.

“We attempted to write our software code in a generalised way from the very beginning. That way, we can use the same software later for various applications,” says Sven Beck, who worked on proximity detection systems for mobile robotics during his Master’s in Robotics, Systems and Control at ETH.

The underlying algorithms on which the software is based came from two Master’s thesis projects at ETH. Together with the experienced ex-Google manager Pedro Marques and their team of ETH and EPFL graduates, the Beck brothers developed these algorithms before then founding a company in autumn 2019.

This sophisticated code is what sets their start-up apart from the competition. After all, in Switzerland alone there are already two other start-ups that have set out to develop automated document recognition solutions. However, no one has yet successfully created a solution that recognises tables – the pinnacle of the field.

Image and text recognition

BLP Digital’s solution uses a combination of two technologies: image and text recognition. Certain algorithms analyse the text, while others scan the document image at the pixel level. “This allows us to position text correctly within the document,” says Sven Beck. Tables can then be recognised and reproduced correctly. In addition, parallel processes create a kind of “double seam”: if their results match, the result is almost certainly correct. “We apply a range of algorithms to the same task in order to eliminate uncertainty,” explains Sven Beck.

BLP could potentially attract customers from any industry that spends a great deal of resources on administrative processes. One challenge for the start-up will be to reach the right markets at the right time. The young entrepreneurs have already started working with the manufacturing industry because they have experience in this sector. The brothers come from a German family of entrepreneurs; their parents own an Enterprise Resource Planning (ERP) company that mainly serves customers in this sector. Before BLP Digital, Tim and Sven had already founded an augmented reality start-up, which is now in other hands.

Early on, the team was engaged in frequent discussions with the industry, resulting in the first pilot projects and sales for BLP Digital. Currently, the company has 10 employees, who work mainly on the first integrations of the company’s software into ERP systems. They receive occasional support from Stefan Feuerriegel, an ETH professor in business informatics. About 50 clients can already process invoices and delivery slips automatically with the software.

Towards the end of the year, the start-up is planning to get investors on board for the first time. Currently, it relies on its own capital and grants from Innosuisse. The additional funds will be used to hire more staff. “Next year, we will be able to automate additional document types for other sectors,” says CEO Tim Beck.