Vasyl Soloshchuk
27 August 2019

Whether it’s Renting the Office or Building AI, the CTO Gets Things Done

Based on the interview with Dr. Alan Bekker, CTO at is a leading product providing human-like user experience for FinTech companies. It uses artificial intelligence, speech and natural language processing (NLP) for customer communication automation in debt collection, insurance, renewals, etc., which allows businesses to scale.

Alan Bekker, a scientist, which recently nominated as one of the top Europe 30 under 30 by Forbes, is the co-founder, and CTO at whose solutions helped the product to win Best of Show Finovate award three times in a row, answered several questions concerning their R&D department routines and his personal motivation.

Alan Co Founded in 2017 together with Mr. Einav Itamar a serial tech entrepreneur which his last company Corrigon was acquired by Ebay in 2016.

Intent is key

Many FinTech startups struggle to maintain numerous customer interactions and communications. Communication with clients often consumes the lion’s share of the personnel’s effort, which could be better spent on creative tasks. Meanwhile, the problem of communication automation is in the machines’ inability to discriminate meaning.

Last year, the NLP and Speech community made a huge advance in understanding natural language. Still, the methods and problems remained the same. A traditional voice bot is just a chatbot that one can plug into a speech-to-text engine from the one side and a text-to-speech engine from the other. Considering the low accuracy rate this approach can provide, this still equals garbage in, garbage out.

“After the first prototyping of, we understood that this approach is not working for a couple of reasons. The first one is that when you are doing speech to text, you’re losing a lot of information on the customer side, for example, intonation. The other reason is that when you do speech to text and you have no context of the conversation, the accuracy rate that you can achieve with great speech recognizers will be still something around 20%, which is still not good enough.”

Before Dr. Bekker, nobody was really thinking about how to create an algorithm that takes the voice of the customer and classifies directly the intent without going through a textual representation. He thought that if he could classify the intent of the customer, this would be a much simpler task than processing speech to text and then classifying the intent. So, Alan and his team built a deep neural network algorithm (called SpeechToIntent) that classifies a voice based on the meaning and intent of the speaker.

“The classical approach for building a voice bot is composed of taking the output transcription from a google API speech recognizer and then feed this as an input for the chatbot engine in order to get the intent, but if you like to classify the intent of the customer, this is a much simpler task than making a speech-to-text and then text classification. The Speech-to-intent accuracy achieved by is near 99%, compared to speech-to-text-intent accuracy, which usually around 70–75% when using Google or Watson API’s, and this makes the whole difference delivering a Human-AI experience as we do.”

Research vs. engineering

The biggest problem of many startups is delivering innovation via continuous research. Usually, the engineering team has a defined path of what and how things should be done. But in data science and machine learning, people usually try to solve research problems where no one knows the best approach. This forces teams and team leaders to think a lot beforehand:

“With continuous research work, you know how to begin the work but not always how [it] will end. A lot of times, you think that the best thing to do is to begin with one approach, but two days after, you discover it isn’t good at all. So, you need to completely change the approach. It’s like you make a plan for two months, but after two days, you need to rearrange it because you just figured out that [the] approach is not working.”

According to Alan, there are three ways to combat the changing environment of a research project. They include having well-defined KPIs, having close communication with the team, and hiring open-minded people who are able to learn on the fly. Here’s what Alan says about KPIs:

“We are always focused on conversations that have well-defined KPI, as it allows making use of reinforcement learning algorithms and maximizes their outcome. For example, if I’m approaching someone from New York, I should use a female voice between the ages of 30 to 35, but if I’m approaching someone from Texas, I should use a male voice [with] some different accent [but] the same wording. You can actually test which words or wordings of the conversation are working better, too.”

Similarly, you need to communicate to the research team what should they do based on results. While managing the data science team, you need to monitor every single step and have daily conversations about whether this or that algorithm is working better on the data or not. 

The third key way to manage these changes is dealing with open-minded, nimble engineers. The cultural fit is important for a startup, so every single person that enters is examined: 

“It means that we need to see that this person is smart enough and has research capabilities. But their personality is even more important. It’s crucial for us to see if this person has the right personality to work in a startup, [if they can work] in a very dynamic environment, and if they’re able to think about solutions on their own.”

Some dreams do come true

Alan likes his job very much, just like he loved his first vocation: coding. When he founded, he thought that his daily work would be just comprise coding. But of course, this is not the reality today. 

“Unfortunately, I don’t have enough time to code by myself. I need to manage the data scientists and developers to do that. My day-to-day routine is split between management and thinking about the technology, trying to analyze the research results, and deciding how to proceed and develop bots that will improve the product. [On] the other side, I have a lot of conversations with existing customers and meetings with investors.”

To make valuable decisions and drive the company to growth, Alan takes part in communities of entrepreneurs and attends meetups where CTOs handle conversations and try to learn from each other. He reads blog posts, listens to podcasts, and tries to use his time carefully to train himself to be a better leader.

“If someone wants to be a CTO or a co-founder of a company, they need to be ready to do everything that is needed for the success of the company. There’s no such a thing [as a] well-defined role in a startup; you need to do everything, from finances, customers, development, and release to just hiring people. So, I would recommend them to prepare yourself mentally and know that the path of an entrepreneur is not well-defined.”

Nevertheless, rolling out a product that can pass intense Turing tests tens of thousands of times a day is a dream for many people around the world. And it came true for Alan and his team; this wakes him up in the morning and motivates all through the day.


Enabling human-like experiences in machine speech conversations is an endless but fascinating mission. So, the role of a CTO in the success of such a startup is huge, and it’s crucial to get things done to achieve unprecedented results.