Apostroph’s Experts Talk About the AI Journey and The Power of Custom Models

Raimon Wintzer and Szymon Ruciński, language technology engineers at Apostroph

Szymon Ruciński and Raimon Wintzer, language technology engineers at Apostroph, share more than just responsibilities at Apostroph’s apoLAB. These digital natives share knowledge and enthusiasm for artificial intelligence (AI) and machine learning. In fact, the two colleagues are passionate about the interaction of these technologies with natural language, and theirs is the kind of critical seminal work taking place in the language services industry and propelling it forward.

Both engineers began getting involved in technology as children, so the current lightning fast AI progression is something they take to naturally and with enthusiasm. Ruciński’s career in IT began with chatbot and voice assistant development, areas he continues to explore at Apostroph. Wintzer has centered his expertise on machine learning, a subfield of artificial intelligence that has become indispensable in the language services industry. 

Their approach to machine learning and AI is two-fold: because they understand the language business, they can look at business processes and at the same time apply a scientific method to arrive at an AI solution that starts with research and data. Every solution for Apostroph’s clients materializes as a result of careful analysis of processes and data, and a healthy dose of creativity.

The Apostroph Answer to ChatGPT

Open AI’s ChatGPT is a powerful model. However, when it comes to specific customer use cases, it may be too imprecise and too general, contend the engineers. Often, even smaller customized models built in collaboration with customers may outperform ChatGPT on domain-specific tasks like medical text translation or finance report classification.

More adaptation and training can truly transform a model to make it suitable for language services, according to Wintzer. He adds that “a refined model specially designed for translation and proofreading is a perfect fit for the range of services at Apostroph.”

Wintzer reasons that multilingual models have a few well-known problems. For example, different languages are in competition against each other within the model for available parameters, and low-resource languages are not well represented in the training data. This means that machine translation still has an opportunity to advance and improve. That is a good thing.

Ruciński further explains that models can be expected to become even better at text comprehension and text generation. Their [artificial] “intelligence” will increase and they will learn to execute complex tasks learning from examples. It is a sort of operation where the model is fed templates from which it can generate completely new constructs. 

We might soon see more end-to-end translation use cases, such as audio-to-text, audio-to-audio, or image-to-text, all without complex workflows consisting of multiple models,” adds Ruciński. The goal is to create simpler yet more sophisticated workflows with models trained on high-quality data.

The Apostroph machine learning experts also agree that models are reaching a level now that enables individuals to accomplish a lot more. Long gone are the days when this type of technology was reserved for technology powerhouses. 

Are you interested in a customized machine translation engine or an AI editing tool hosted in Switzerland? Let’s talk! Contact Nadia Gaille, Business Success Manager, sales@apostrophgroup.ch, 41 44 265 40 30

Better Data Security is Also Here

Data storage vulnerabilities have been a constant concern for many in the business world, and it is not different in the language services industry. Data security is and should always be a factor for b2b collaboration, and at Apostroph, explained Wintzer, data security is constantly evolving so that the company stays ahead of any risks and clients are reassured that their own data is safe and secure.  

Furthermore, Apostroph continues to increase its in-house storage solution to ensure all data and related systems remain in Switzerland at all times. Hardware capacity is also increasing so that, as the company implements local, custom language models, even in-house development relies less and less on out of country clouds.

Image: Raimon Wintzer (L) and Szymon Ruciński (R), language technology engineers at Apostroph Group