• About DIVISIO
  • AI Basics
  • Amazon DJL - a new DL framework for Java
  • BGL symposium 2019 - lecture 'AI and Magic'
  • Categories
  • ChatGPT
  • ChatGPT for Teams: Privacy-Compliant Use in the Workplace
  • Conferences
  • Contact
  • CUDA
  • Deep Fakes - How to spot faked Images
  • Deep Java Learning Introduction - Part 1: NDManager & NDArray
  • Deep Learning
  • Demos
  • Depixelizer
  • DIVISIO Blog | Understanding AI, DL and ML together
  • DJL
  • DL4J Workshop at the ML Summit in Berlin
  • Enterprise Tensorflow - Java vs. Python
  • Enterprise TensorFlow 2 - Saving a trained model
  • Enterprise TensorFlow 3 - Loading a SavedModel in Java
  • Enterprise TensorFlow 4 - Executing a TensorFlow Session in Java
  • Enterprise Tensorflow: Code Examples
  • Git as a management tool for training data and experiments in ML
  • GPT-3
  • GPT-4
  • Homepage
  • Imprint
  • JAVA
  • Jax 2018 - Talks about DL4J and more
  • Jax 2019 Recap
  • KI Basics
  • KI Basics"
  • Konferenzen
  • ML Conference 2017 in Berlin
  • ML Development
  • MLOps: Establishment and operation of an AI
  • Neural networks - The five most common mistakes
  • NLP
  • NLP, NLU and NLG: AI and text
  • OpenAI API
  • Privacy Policy
  • Python
  • Python Dependency Management with Fewer Headaches
  • PyTorch
  • Recap: ML Conference 2019 in Munich
  • References
  • Sokratext NER - Analyze German texts using Java
  • Sokratext NLG - Generate thousands of German texts with just one click
  • Sokratext POS Tagger - Automated Word Type Recognition with Java
  • Sokratexts
  • STAG AI Image Tagger - Bilder lokal mit KI taggen
  • Tags
  • TensorFlow
  • TensorFlow and Java - An interview with entwickler.de
  • Types of Artificial Neural Networks
  • Understanding AI - Part 1: What is AI?
  • Understanding AI - Part 2: Symbolic AI, Neural Networks and Deep Learning
  • Understanding AI - Part 3: Methods of symbolic AI
  • Understanding AI - Part 4: The basics of Machine Learning
  • Understanding AI - Part 5: Supervised & Unsupervised Learning in ML
  • What are Neural Networks and how do they work?
  • Whisper 3 Large for JAVA
  • Working with Ollama, Part 1
  • Working with Ollama, Part 2
en

Customer References

DHL

DHL is the leading provider of identification solutions in Germany, offering services in branches, at front doors, and online via video. However, with massive improvements in computer vision through deep learning, even complex processes like identity verification can be fully automated.

For this purpose, we developed the new AutoIdent procedure for DHL, which is now available to DHL’s business customers alongside proven methods. This involves comprehensive analysis of identification documents, extracting the document type and all information. The key feature: the procedure was developed with data efficiency, so new documents can be supported with just a single example image (e.g., the sample ID from the Federal Printing Office) and a configuration file. This is useful for fully automatic verification and can also save valuable processing time for agents during manual verification.

In further steps, a series of identity and security features are verified. Overall, more than 10 neural networks are used for this purpose. These were selected and trained by us and integrated into DHL’s infrastructure.

The key advantage is that this is a full-featured JAVA Enterprise solution, so operation of the new system seamlessly integrates into existing infrastructure, code guidelines, and quality requirements.

Unlike competing solutions, our solution for DHL is real-time capable, meaning the artificial intelligence systems don’t process pre-recorded videos but can interact with end users, e.g., providing process instructions. A new WebRTC interface makes the system future-proof, especially in the age of generative AI, and enables it to eventually take over all tasks of a human agent.

onvista

onvista media is the leading provider of news in the retail securities market. Through the well-known portal onvista.de, it offers high-quality editorial content alongside stock prices and hard data. However, due to the abundance of information, even onvista’s professionals cannot cover all news, especially in the long-tail area.

Therefore, long before the ChatGPT hype, we developed text generation for onvista that creates text summaries for securities based on onvista’s price and securities database - securities that don’t justify the effort of human editors due to their low trading volume. This way, content can be generated for the extensive long tail of the securities market that still provides value to interested readers.

This system includes a self-developed transformer for German language generation and an automatic analysis system for financial data that generates input for text generation in a deterministic and traceable manner. The entire system is configurable through an elaborate Vue.JS interface and can generate hundreds of texts per second.

Goldschmidt

The Goldschmidt Group has been the expert in rail infrastructure for over 125 years. In addition to thermite welding, invented by Hans Goldschmidt in 1895, Goldschmidt is also an expert in quality testing of rail infrastructure.

For this purpose, Goldschmidt has developed an ultrasonic measurement system that enables non-destructive testing of rail networks on inspection trains at speeds of 70 km/h. This system is in use at Deutsche Bahn, among others. The results of these measurements are currently manually evaluated by trained specialists, who then initiate necessary further controls and maintenance measures.

Together with the measurement technology and infrastructure experts from Goldschmidt and Deutsche Bahn, we developed a deep learning procedure that will improve evaluation quality in cooperation with humans while significantly increasing evaluation speed, so that more measurement trips can be analyzed in the same time - leading to significant quality and efficiency improvements in rail maintenance.

This project employs state-of-the-art analysis techniques, supervised and unsupervised learning. The system is specifically developed for Goldschmidt’s measurement systems and trained from scratch on terabytes of data.

For security reasons, the finished system runs directly on the customer’s infrastructure and is not dependent on non-European cloud providers.

REWE

REWE is not only one of Germany’s leading retailers but also a pioneer in digitalization and technical innovation in retail.

For REWE, we successfully supported a demanding project in computer vision. Our experienced AI team helped the company’s internal developers expand their deep learning competencies and implement a customized system that significantly increases the efficiency of internal processes.

In close collaboration with the customer’s dedicated development team, we contributed our expertise in data acquisition and management, building a powerful training infrastructure, and developing and testing optimized training scripts. A particular focus was on seamlessly integrating the solution into the customer’s existing on-premise infrastructure to ensure secure and reliable operation without dependence on internet connections.

Through our consulting and guidance, the retailer’s internal team could significantly expand their deep learning capabilities and successfully implement the project independently. We are proud to have contributed to expanding existing know-how and accompanying the company on its path to a forward-looking, AI-supported solution that creates sustainable competitive advantages.

Continental

Continental stands worldwide for innovative tires of the highest quality. Therefore, we are pleased to have supported a forward-looking project in non-destructive material testing there. Our expert team helped the company’s engineers integrate innovative approaches of semi-supervised and unsupervised learning into their testing processes, elevating quality control to a new level.

In intensive on-site workshops, our AI specialists worked closely with the tire manufacturer’s specialists to develop customized solutions that seamlessly integrate into the company’s existing Java infrastructure. Together, we developed concepts for data collection, processing, and analysis that enable early and reliable detection of potential material defects without impacting production processes.

Through knowledge transfer and hands-on guidance from our team, the tire manufacturer’s internal experts were enabled to independently develop the novel testing methods further and extend them to other application areas. The successful implementation of AI-supported solutions helps continuously improve product quality, minimize reject rates, and secure the company’s competitiveness in the long term.

We are proud to have contributed to expanding the tire manufacturer’s know-how in semi-supervised and unsupervised learning and jointly developing forward-looking solutions that create real value for the company.

UNICEPTA

UNICEPTA is one of the leading service providers for Media Intelligence, not only in Germany but also worldwide. Today it supports companies and organizations in Europe, North America, South America, and the Asia-Pacific region. In addition to an expert team for media analysis and selection, UNICEPTA increasingly uses AI technologies for its work.

In this context, we were able to develop an embeddings-based text classification that can automatically create press reviews for each customer individually based on previous human selections. This means that after a few examples from human experts, the system can identify relevant articles for a customer, massively simplifying the creation of new press reviews.

Here too, transformer models were used. The technologies used are the same as those used to develop RAG systems.