September 26

“Building intelligent machines”
by Claudio Huyskens,



Who are you and what do you do?

I am 41 years old with an educational  background in management – specifically technology management (PhD). About 15 years ago, I discovered my interest in software development andI went from curiosity to financing my studies as a freelance software developer. Inspired by quantitative research for my PhD, I have since 2008 followed that passion in different fields of application, ranging from quantitative finance to machine learning and artificial intelligence. I am  the founder and CEO of, a Cologne-based startup providing Machine learning and Artificial intelligence services to businesses.

Why did you become a startup founder?  

In retrospective, the first touch point with an impact on my personal development in that direction was back in 1994 during a career day in my school in Cleveland, OH. Of all tracks, the entrepreneurial track struck me most, as the narrative of the advocating entrepreneurs was full of initiatives and creativity. That experience probably influenced me,  that was when I decided to rather work self-employed to finance my studies.

My PhD advisor and her mantra ‘Do it, don’t just talk about it!’ certainly also motivated me in that direction. I believe many rounds of job interviews with consultancies and media / technological corporates in 2008 probably brought a final confirmation and guided  me towards becoming an entrepreneur. Together with my fedger co-founder, Benedikt, I went through a trial and error phase of founding and shutting down twice, before we started fedger in 2014.

How did you come up with the idea for Fedger?

I must admit, there is no surviving ‘founding idea’ in the solution or product sense. I will say we rather built fedger on a broader vision of data accumulation and analysis for better and eventually automated decision making. Starting without a clear product idea definitely made life tougher during the first two years, but it taught us a couple of extremely valuable lessons such as ‘if you cannot pinpoint what pain your solution helps to relieve in one sentence, you can’t build a business on it’.

I believe that establishing a defendable startup idea is extremely difficult and rarely seen among early stage startups. In that sense, our painful experience definitely sharpened my focus on clarity of idea. It’s important, though, not to confuse the criticality of the idea component with the fact that idea will also undergo an evolutionary process; any feedback, any experience and any change of the environment needs to be reflected in the continuous revaluation and re-articulation of the business idea.

The product you are offering is a fairly technical product. How would you describe it to someone that doesn’t have a technical background?

Customers use fedger AI technology to automate information extraction tasks. In practice, delivery services or POS vendors for example use the fedger menu extraction API in their customer onboarding process to automatically extract the transactional information content from a given restaurant menu card. The service offers them greater flexibility, vastly reduced runtimes at lower cost. As input, the service accepts PDFs or images of restaurant menu cards and returns a structured CSV or JSON file containing categorized items, descriptions, quantities and prices. The service largely automates a process, which used to require manual copy typing. Most people think, the invention of OCR had solved such challenges, but that is not valid for structured documents such as forms or tables where the appropriate semantic type of an information such as price or quantity are necessary elements. Fedger technology works similar to object detection that recognize objects such as a car in a photo.  

Who are the founders of Fedger and how did you find each other?

I co-founded fedger with my friend Benedikt Knobloch. I met Benedikt back at the University about 15 years ago. We have since then studied together, worked together in academic research, we have even shared an apartment for a short time. We have over time often heard the discussion on founding team composition and I fully agree with the US venture capital research arguing that – beyond complementary skill sets – nature and length of personal relationship is detrimental to startup success.

Since founding fedger in 2014, we have been on (and sometimes over) the edge alot and in a startup crisis had an impact far beyond job and professional life. As a founder, you are ‘all in’ and crisis means survival mode. Only a joint track record of a trust relationship helps  in keeping things together to jointly spearhead in the same direction.

What are your main responsibilities at Fedger?

My role at fedger has undergone some transition over the last two years. I started as CEO / CTO, responsible for product development both conceptually and technically. 18 months ago, we  employed a great CTO, to take the technical development and infrastructure to the next level. This also freed capacity on my side to focus more on business and product strategy.

Today, I see three main fields of responsibilities in my role: First, extending the core technology to service lines in new market segments as well as driving our core service line deeper into the value chain of our customers. Second, developing the organization to manage growth including process definition, responsibility assignment and partnerships. Third, manifesting a company’s culture and value system serving as DNA, inspiration and reference for new team members and new challenges.

How is Fedger financed?

We raised two smaller angel rounds in 2014 and 2016 for a total of about 100k and are currently cash flow positive.

How many people in the team work in tech and how many in business?

We currently have a 2:1 ratio between tech and business with 8 fulltime engineers and developers and 4 fulltime business managers. In addition, we have 16 working students, of which 2 support the engineering, 4 support business management and 10 work in ops, i.e. service execution and delivery.

What tech stack do you guys work with?

Language-wise, we try to limit any development to Python. We work with Keras and TensorFlow as Machine Learning development environment and we use Docker containerization and Kubernetes for their orchestration in Google Cloud.

For API management, we rely on Kong. Besides, we are largely agnostic and try on a case by case basis to select technology primarily according to task requirements.

Where do you apply AI technology? What are the technological challenges in applying AI?

We have developed an AI for information extraction and have first applied it to retrieve the items from restaurant menu cards. We currently evaluate applications to invoices and insurance documents. In all those cases, fedger AI solves the problem of document structure recognition and resolution. The larger the structural variance among documents, the greater the benefits from AI.

Having sufficient amounts of training data available marks the greatest technical and procedural challenge in the development of such an AI. Actually, technology is the least among the challenges with regard to AI. Crossing the border from deterministic to probabilistic means a change of paradigm in many senses. For instance, regulatory bodies such as financial service regulation requires a comprehensive functional description of how data input is transformed and what output can be expected – deterministically.

The fact that AI solutions implicitly involve training towards global optimization means applications do not necessarily create equally distributed benefits, in other words, AI discriminates and we may not be able to explain the extreme outcomes. As opposed to opportunistic application development from academic AI research, developing an AI for a certain application purpose brings along what I call the “AI investment paradoxon”. Be it in a corporate or startup context, any funding requires a detailed plan including budgets to be spent and a milestone roadmap outlining expected delivery dates and qualities. Comparable to bringing up a child, AI maturity and problem solution capability cannot be planned precisely, which thereby makes investment decisions by current standards impossible.       

How do you currently measure success? What core metrics do you look at?

We measure three dimensions of success, that is, market success, delivery success and developmental success.

Regarding market success, we focus on the number of monthly service subscriptions and even more importantly on the number of monthly API transactions. To us, the ladder KPI provides the best integration of customer growth and growth of customer transactions.

Regarding delivery success of our Artificial/Human Intelligence blended service, we focus on transactional KPIs measured against our asynchronous API. Such metrics include round trip duration and net labeling time per transaction respectively. We also track task complexity to determine proper sampling and benchmarking. Towards empirical AI evaluation and assessment, we track the above mentioned runtime metrics across a broader sample over time.

Regarding developmental success, we have implemented pretty tough testing requirements throughout, which basically takes out the quality component and reduces success to productivity measurement. For that purpose we use a typical agile story point logic.

What were your biggest stumbling blocks when you started?

Two main hurdles stuck out: funding and – for a B2B startup this is  extremely important – credibility. This posed a chicken-egg-problem insofar, as reference cases typically provide the credibility you need to convince early customers. Running free-of-charge proof-of-concept projects to provide such a reference, required funding. To even get consideration for a funding, though, Institutional investors required market traction or at least successfully conducted MVPs. We resolved that deadlock, when we convinced Alexander Marten, a startup founder himself, to commit to an Angel round. In retrospective, that investment was clearly idealistic and high-risk.  

What are the biggest challenges you are facing today?

Growing and managing development capacity under financial resource constraints.

What advice would you give founders to be?

Go out, be open minded, contribute and above all, listen to mentors i.e. people willing to help you with their experience without necessarily benefitting from it.

AI is an exciting technology, which could bring a new wave of significant technological breakthroughs. Some people like Elon Musk warn about the dangers of AI. In which camp are you?

Personally, I do not perceive intelligent agents as danger or risk. Many of the end of world scenarios, we can read about in tabloids today are total nonsense and reflect one thing:” It is still a lot easier to scare people away from a development than it is to educate and raise curiosity for the inner workings of a technology and the great potential it offers to solve some of the seemingly unsolvable man-made problems, we are facing today”.

Also, I just don’t see the reasoning of a truly intelligent agent to harm mankind. Many of the potential motivations and behaviors described in the press are specifically characteristic of human psychology and sociology and have long been proven to have logically more superior alternatives. So, taking it all into consideration, I am very optimistic about the coexistence of man and machines.    

Cologne is for Startups…

Cologne is a great, authentic startup hub, offering great and affordable talent, a super open-minded atmosphere and a range of experienced and extremely supportive mentors and angel investors. Early on, we considered moving fedger abroad. As we started growing the business, the benefits of the Cologne/Bonn area played out really well for us.


You may also like

{"email":"Email address invalid","url":"Website address invalid","required":"Required field missing"}

Subscribe to our newsletter now!