EXPAI offers a tool that allows you to understand the reasoning of AI models. We believe transparency adds value at several levels within a company. Data scientists will have a tool that will allow them to accelerate and improve the assessment of their AI models, identify errors and enhance the transfer to users with less technical profiles. And once EXPAI has been integrated into the company’s AI development process, the business leaders will be able to better understand the decisions made based on its predictions and increase the benefits yielded by these technologies. Finally, end users, the people who consume the AI predictions, such as sales associates, will have much more information to work with when making their own decisions. Amongst other things, this will make it possible to detect errors in real time, expedite processes and build confidence in this technology.

EXPAI offers a tool that allows you to understand the reasoning of AI models. We believe transparency adds value at several levels within a company

Javier Rando, one of the start-up’s founders and a UPF alumnus, had previously worked with the university’s research group specialized in identifying and mitigating the dangers of AI. This field is known as algorithmic fairness. Some of the best-known scandals are the recruitment algorithm that Amazon had to scrap because it was sexist or the tool used to help US judges decide whether to keep people awaiting trial in custody that was biased against African Americans.

The team’s founders also have experience working with AI in business, which has brought them face to face with some of its limitations. One example we always cite is the creation of a system for identifying customers likely to leave a company. After months of development, the sales team tossed it because they did not understand its predictions.

In our view, there are several dangers. First, there is the possibility that an algorithm will work better for some groups from our data than others. If these groups are determined by personal characteristics such as ethnicity or gender, then the result will be discrimination. Even if we are not working with people, such differences in performance can lead to major mistakes and suboptimal decisions. Second, the opacity of AI models can make end users and customers reluctant to use them, which jeopardizes the development of a company’s operations.

AI models’ opacity can make end users and customers reluctant to use them, which jeopardizes a company’s operations

When data scientists create an algorithm (or model), they can choose to make a very simple one (e.g. a regression model or decision tree) or an extremely complex one (e.g. a neural network), depending on the task to be solved. These models will learn to replicate a behaviour based on training data. For example, let’s say we know all the customers who have left or stayed with our company in the last five years. We can use that information to train an AI algorithm to predict which of our current customers are most likely to leave. In this training process, the algorithm adjusts its internal parameters to learn patterns in the data without anyone specifying exactly what it should do. In other words, it does this without anyone telling it prior rules, simply by using the data.

Now, data scientists have many ways to measure whether a model is returning appropriate results, but in the end, the operations being performed within the algorithm are so complex that it is impossible for a human being to understand the ‘reasoning’ taking place within it. All they are going to see is the result, i.e. whether or not the customer will leave the company. Our goal is to accompany these predictions with the reasoning the model has followed, in order to build trust and speed up decision-making.

Our work focuses on making models transparent by creating a bridge between academia and industry. We choose the best explainability techniques developed in academia and implement them in a way that is easy for industry to consume and understand. These techniques are capable of extracting the reasoning used by the model at different levels to provide insight into how it works on a global scale and into each of its predictions.

At EXPAI, we do not work directly on the assessment and detection of biases. We offer various tools to help companies detect disparate performance between data sets during development.

Like many emerging technologies, AI raises some concerns amongst the general population. One key factor is the difficulty of understanding how it works and even of knowing exactly what it is and what it is good for. In fact, for a lot of people, the first thing that comes to mind when they think about AI is a sci-fi film about an intelligent robot that has put humankind at risk.

For a lot of people, the first thing that comes to mind when they think about AI is a sci-fi film about an intelligent robot that has put humankind at risk. However, in reality, we are still far removed from that

In reality, though, we are still far removed from that type of intelligence. What many people do not know is that many simpler AI systems are already used in everyday life to make our lives a little easier.

At EXPAI, we are optimistic. We believe that a responsible use of AI that ensures its transparency can have a very positive impact on our lives.

Having people with different academic backgrounds can only enrich any business. On the one hand, we have the technical team, which is mainly responsible for product development and consists of us, both of whom are graduates of the Mathematical Engineering in Data Science programme. On the other, we have Carles Soler, an adjunct lecturer at UPF, who is in charge of the company’s operational and commercial affairs.

Even though we started our project outside the university, from the outset we have received the support of a wonderful team that works to increase the university community’s chances of success with regard to our entrepreneurial activity. Being part of the ‘In residence in the Tallers area’ programme has been a great help. For one thing, it has provided us with a space where can meet and be in contact with other companies in situations similar to ours. It has given us a room to use as our own to have a physical place to meet and go to work, as well as a lot of external resources to get EXPAI off the ground.

Participating in 4YFN was another major opportunity that UPF afforded us. Attending such an important event was a great chance to get to know the ecosystem and get the word out about our value proposition. The response was really good and we managed to establish contact with major companies.

Our product is currently on the market and we are in the process of securing our first clients. EXPAI has been well received by the market. As a result of an impactful marketing campaign in the media and the award we won as a finalist in the AI Research and Development category at the recent edition of Advanced Factories, some ten companies are currently testing the value our operational service provides with free trials.

Predicting the future is always uncertain in the world of start-ups. Right now, our goal is to get our first clients, which will allow us to expand our team in order to continue improving our solution organically.

Entrepreneurship is definitely a complicated path. The founding team has been working without pay for more than a year, with the conviction that it is going to work. Some of us are also still studying and working, which means sacrificing some of our free time as well. And all of that is compounded by human factors, such as changes in the team, personal problems, etc.

But the overall experience has been very positive, which is why we are still here today, excited to keep learning and growing.