Since the beginning of 2018, the 26 software specialists at Malmø-based Neodev have been on an adventurous journey. Back then CEO Fredrika Olsson and her team decided to focus heavily on bringing Machine Learning technologies to Swedish businesses and public sector organisations. The gamble has paid off. Neodev has now established itself as a Machine Learning expert well-known in the IT sector of Southern Sweden, with currently half of its workforce engaged in Machine Learning projects with a wide range of clients.
Along the way Fredrika Olsson has learned some important lessons from her efforts to bring Machine Learning from the fluffy world of IT buzzwords to real-life applications. Maybe the most important one – as mentioned initially – is the fact that actual Machine Learning algorithms only constitute one fifth of the work you must put into a project. The bulk of the work is data and deployment.
What data do you have?
- When we do workshops with clients to help them decide if they should integrate Machine Learning into an application, the biggest question is always “What data do you have?”. If you don’t have any data, you can’t do anything. If you have enough data, there comes another problem. Often it is saved in different formats, and merging it requires a lot of work. Also, you must clean the data to reach the quality required to make it fit into the Machine Learning environment. These are the most common challenges we see.
- And on top of that, deployment is a huge issue as well. When you’ve finished building your application you must have it running somewhere, for instance in the cloud. So to sum it up: data is 40% of the work, deployment is another 40%, and that leaves 20% for working on Machine Learning models.
Tweaking existing models
Fredrika Olsson emphasises, that Neodev’s engineers are not in the business of developing Machine Learning models themselves. That is done by university researchers and very large companies. Instead, based on their understanding of existing models, they tweak them to make them work for specific use cases.
That expertise has attracted a lot of projects and assignments since it all began in 2018. Neodev is both doing in-house proof-of-concepts and end-to-end development projects, as well as providing Machine Learning consultants to customers’ development teams.
- And it all began with the decision to back in 2018 to dig into Machine Learning. We had two reasons for it. Firstly, a lot of our developers were interested in it. Secondly, by that time the technology had become so good you could begin to do things you hadn’t been able to before.
The beginning: A course
To present the technology, Neodev developed a two-day introductory Machine Learning course targeted at experienced software developers from other companies. As Neodev is a software development company and not a course provider, this was a one-off. However, the approach soon proved to be a success, and by the beginning of 2020 the course had been run 15 times.
Then corona hit and suddenly the demand for face-to-face training came to a complete standstill.
But by that time the many courses had created such a demand in the market, that Neodev had their hands full with doing workshops, proof-of-concepts, and complete development projects, centred on Machine Learning.
And the business has grown ever since.
- We’re doing a wide range of projects. Some of them are financed by the customers themselves. Others are partially funded by the government, as Sweden has a national ambition of becoming one of the world’s top countries in AI and Machine Learning.
It is well-known, that computer vision is the maybe most developed application area when it comes to Machine Learning. Therefore, it comes as no surprise, that many of the projects Neodev is working on are in that area.
For instance, Neodev has helped a large company, manufacturing camera systems for home security, prevent false alarms. As many alarms are triggered by movement sensors activated by animals, a model was developed to enable the system to distinguish between people and pets. As a result, preventing false alarms lowers security companies’ cost of providing services for their customers.
Another computer vision application was built for a start-up selling digital services to interior designers. The company named Boni provides a platform for designing the interior of rooms and offices digitally, with images of furniture provided by manufacturers. However, many of these images show the furniture on white background, while the Boni platform needs a transparent background for seamless integration. Neodev developed software, based on a pre-trained object detection model, that automatically changes white backgrounds to transparent ones, without erasing parts of the products or keeping parts of the background.
Large Language Models
Neodevs Machine Learning team has also developed applications based on Large Language Models (LLM). For instance, a company providing purchasing systems for large public administrations required a tool for the categorization of purchases. The goal was to be able to produce reliable purchasing analyses to create a base for choosing the right purchasing strategies. An AI-based system was developed for categorization, to identify products and services in one’s purchase. The overview created makes it easier to analyse complex purchases in large organizations, among other things establishing a common language between finance and purchasing departments.
For this, a team of three Neodev developers built and implemented a tool based on a NLP Bert model, with a further improved LLM based model to be put into operation shortly.
Also, Neodev has designed and implemented a decision support tool for improved allocation of delivery capacity. Experiencing issues managing the stocks, involving either storage or goods shortage, the client needed a tool for visualization of the current situation in terms of stock and storage space. Neodev developed a solution, also including forecasts and prediction models to highlight upcoming problem spots.
The right decision
Looking back across the five years that have passed since deciding to place a large bet on Machine learning, Fredrika Olsson states that much has happened.
- It has become possible to do so many things. Computers have become smaller, cheaper, and more powerful. Five years ago, some of what we do now would only be possible in a research center at Google. Now a lot of companies can afford to implement Machine Learning technology and create a viable business case around it.
- One of the new things we’re going to see will be embedded devices with quite powerful Machine Learning possibilities. Already you can do Machine Learning calculations on small devices. That trend will become much stronger, and it will open even more opportunities for expert consultants like us. No doubt, back in 2018 we made the right decision.