STRG’s Blue-Sky Research

STRG’s Blue-Sky Research

Our product is our expertise

It is expert knowledge that springs from intensive research, as well as from human and machine learning. We may refer to STRG.CMS or STRG.BeHave as our “products” and we may call some of our Agile co-workers “Product Owners,” but in reality, STRG produces nothing it can stuff in a shrink-wrapped package and sell over a counter. We offer no downloadable apps nor white-label web services. Nor do we publish any pricelist of our standard fees or package rates.

What do we do, then? A $#!+load of research, that’s what! Our research doesn’t have the goal of creating a finished product. Rather, it yields constantly evolving toolkits that STRG’s talented teams use to solve clients’ problems.

STRG is a tech-driven enterprise. Whenever we encounter a challenge that current technology can’t solve, we go into research mode.

From concept to practical tools

One of the earliest applications of our research efforts was STRG.CMS, which was one of the very first content-management systems to utilize semantic-analysis techniques. But the existing tools for creating personalized news content burdened editors with the task of tagging, categorizing and labeling content. We saw this as a challenge that could be solved with machine-learning methods.

Shortly after Google published TensorFlow in 2016, STRG’s CEO Jürgen Schmidt came up with the idea of initiating a research project to dig deeper into the semantic analysis field, including Natural Language Processing (NLP), word embeddings, and supervised-learning methods. Initially, he intended the research to result in a product, at least one for internal use. In 2018, perfect funding – neither strictly for research nor product development – was granted by the Austrian Research Promotion Agency (FFG). Research was carried out over two years and the outcome was STRG.BeHave.

“Instead of developing a marketable product for others, we developed proprietary toolkits that we use internally to develop applications or e-commerce portals,” says Eugen Lindorfer, a eight-year veteran Product Owner at STRG. “We discovered that we are not a product company, but a project company. Everyone says you can’t do both, because each requires a completely different organizational structure and mode of thinking. Once we changed our viewpoint from being a product developer to being a problem-solving project company, our business really took off.”

“We are not a product company, but a project company. Everyone says you can’t do both, because each requires a completely different organizational structure and mode of thinking.” 

While other tech-driven entrepreneurs are always looking to develop the next killer app, STRG has avoided the temptation to develop plug-and-play software solutions, and has instead marketed its research-based expertise and know-how. “Of course, we could have produced something similar to Outbrain,” says Lindorfer, “but this is only a small part of what can be done with BeHave. If we launched BeHave as a product, potential clients would just compare it with Outbrain, which I see as ineffective, especially in terms of enabling customers to monetize their content. Plus, I don’t think we would earn much revenue from such a product.”

One of the first clients to benefit from the FFG-funded research on BeHave was the Austrian weekly newspaper, Die Furche. While helping them to design and develop their online portal, we discovered we could apply BeHave’s semantic Natural Language Processing (NLP) toolkit to organize and display their massive archival content dating back to their founding in 1945. 

“We developed a timeline feature, which they named ‘Navigator,’” says Lindorfer. “It uses AI-enhanced semantic analysis to automatically categorize and tag both contemporary digital content and their older, digitally scanned content – even the poetry, which they stopped publishing over 50 years ago.” The elegant timeline-slider function enhances the reader’s experience by automatically placing contemporary content into historical context.

Die Furche has its own talented IT department that might have opted for some plug-and-play app to enable display of related content, but this would have required untold hours of editorial effort to label and categorize their archives manually. “We got the project not because BeHave was an off-the-shelf product we could sell to them,” believes Lindorfer, “but rather because it was a tool allowing us to tailor their technology to their specific requirements.”

Beyond the news

Thanks to its research and development of STRG.CMS and BeHave, STRG has been able to secure a strong reputation working with clients in the digital news media sector. As we conducted research on BeHave for this market segment, we realized there was potential for applying its various toolkits and packages to diverse fields, such as ÖAMTC.

“When we do research, we sometimes find solutions for issues that we were not trying to solve,” explains Lindorfer. “By not focusing on a specific product as the outcome, our reputation as tech-driven researchers attracts the attention of many clients from different fields who want a bespoke solution for their portal.”

“By not focusing on a specific product as the outcome, our reputation as tech-driven researchers attracts the attention of many clients from different fields who want a bespoke solution for their portal.”

Agents for change

While research on BeHave is ongoing, it has evolved into a new FFG-funded research project which we are calling “STRG.Agents.” The idea arose (as most usually do) from trying to solve a problem. One of Austria’s largest mail-order/e-commerce distributors wanted to enhance its site’s personalization features with the kind of product recommendations and automated user-interface technology used by so many e-commerce portals, including Amazon, with mixed results (have you ever wondered why you are fed product recommendations for refrigerators even after you just bought a new one?).

“We wanted to improve on this using the semantic analysis methodology we developed with BeHave,” recalls Lindorfer. “We tried to use BeHave algorithms to semantically analyze short product descriptions and related them with strongly structured product data. At first it didn’t work too well, but then we discovered why. The site, while highly popular in Austria, just didn’t have the traffic volume to create a suitable data set. What could we do about that? How could we produce insights into websites that don’t even exist? These two questions were the incentive behind our next research project.”

After many wine and coffee-fueled discussions, STRG.Agents was set into motion this year. The ‘Aha!’ moment came when we realized that we could gain insights for e-commerce platforms by analyzing user behavior on news-media portals – not semantically but how users interface with the site. However, stringent GDPR user-privacy rules prohibit using individual user data for unintended purposes. Nevertheless, Lindorfer says, “We realized it is possible to simulate site visitation using virtual users, whom we call ‘Agents.’” Agents is a term used in Reinforcement Learning, an AI technology enabling machines to autonomously learn within any environment, similarly to how a puppy learns new tricks – a combination of actions, rewards and observation.

“We realized it is possible to simulate site visitation with virtual ‘Agents,’ which could make up for a lack of significant real-user data from smaller web portals or even predict traffic for web-pages not yet built.”

Lindorfer and STRG’s data scientist believe that web-portals can be modelled as a directed graph, and user interactions in a web-portal can be represented as a Markov reward process containing states (i.e., what web-page one is currently browsing), actions (where do I scroll to or click on) and rewards (based on a limited data set of real-user conversions) to determine probabilities of moving from state to state. Running such a virtual-user simulation could make up for a lack of significant real-user data from smaller web portals or even predict traffic for web-pages not yet built.

“We have yet to apply our Agents research to specific applications,” admits Lindorfer, “but we plan to use it for optimizing e-commerce portals, not only to improve product recommendations but to improve the overall UI, in terms of where to place links, how big they should be, and so on.”

The research project is divided into several work packages and FFG funding has been granted for one year. “Ultimately, we want to be able to create a graph representation of a web-portal automatically by simply typing in an URL,” dreams Lindorfer. “But we also want to model pages that don’t yet exist, as well as to create a drag-and-drop back office system to customize learning structures and gain insights into the results of Agent simulations.”

Harvesting the fruit of our research

Of course, constantly evolving technology can render any long-term development project obsolete before it’s completed, so Agents research will be applied on the fly. STRG CEO, Jürgen Schmidt, says “you can’t wait until something is fully ripe before taking it to market. Agents will need at least two years of development, but already in 2022, we will start applying some modules to our current projects.”

STRG’s research can’t just focus on near-term goals. “A company seeking software that automatically classified images approached us for a custom solution,” recalls Lindorfer, “but before we could seal the deal, they found some new, off-the-shelf, open source solution that perfectly suited their requirements.”

Though Lindorfer believes it’s unlikely that reinforcement learning will be replaced by something else anytime soon, neither he nor STRG’s data scientists can always spare the time to keep up on all the current and future developments. “Within this field the algorithms and libraries are constantly in flux. That’s why we have a partnership with the Austrian polytechnic college, FH St. Pölten. These young academics have the time to attend all these super-nerdy conferences and keep us up to date on the cutting edge. For example, we typically use TensorFlow and PyTorch frameworks for implementing Machine Learning, however a researcher at St. Pölten learned at a conference that Google just released the JAX framework.”

For now, Facebook and Google are making some machine-learning software libraries open source. But if they’re allowed to monopolize the market, you know they’ll end up monetizing it. “We can’t leave this research up to these large companies,” warns Lindorfer, “or we’ll become too dependent on them. They aren’t looking to make the world a better place – their mission is to dominate the market and make tons of money.”

By focusing on blue-sky research instead of on product development, the insights STRG gains can always be applied to something completely different from the original goal. It will never be a wasted effort.

If issues with smaller e-commerce companies can be solved using STRG.Agents research, the same technology could potentially be applied to solve larger societal issues like clean energy and the global supply-chain, as well as help to level the playing field between the global commerce monopolies and local SMEs.

To find out exactly how STRG’s research can help your existing digital business or plan your digital debut, drop us a line!

Call STRG’s CEO Jürgen, or write a mail juergen.schmidt@strg.at

Mobile: +43 699 1 7777 165