Artificial intelligence

We integrate artificial intelligence for data-based decisions

The targeted use of AI technologies creates new opportunities to make data-based decisions, make processes more efficient and act with foresight. This makes your company not only more agile and automated, but also more sustainable in terms of its competitiveness.

Use cases

A selection of our work

Read out invoice data with LLM

Initial situation​

  • Companies receive many invoices in different formats and designs every day
  • Classic OCR or rule-based systems reach their limits, especially with unstructured PDFs or scans
  • Manual recording of invoices is time-consuming and error-prone
  • In most cases, horseradish system integrated models are not as error-free as established LLMs

 

Description 

  • Evaluation of an intelligent language model that best meets the client's needs
  • Preparation for extracting relevant invoice information directly from the flow text
  • The LLM is controlled via an API and returns structured outputs
  • Connection of peripheral systems for automatic further processing of data without media discontinuity using Power Automate

 

Added value / goal

  • Reduction of manual work and susceptibility to errors during data entry
  • Processing even with inconsistent layouts, language variants or low-quality scans
  • Automatic forwarding or validation based on extracted information (e.g. assignment to contracts)
  • Scalable solution for SMEs and large companies (integration with peripheral systems e.g. Powerplatform)

Intelligent chatbots

​Initial situation

  • Numerous pages of business documents are generated every day, in PDF and other formats
  • Nobody has the time to read through them, but they are time-consuming to create
  • If someone is looking for information, then he knows where to look or spends a lot of time until he finds it.
  • The documents are updated on a recurring basis, long-term employees do it the way they have always done it and do not look at any adjustments

Description ​

  • Definition of the source for a specific chatbot
  • Building a chatbot with the customer's specific functions
  • Definition regarding the use of the chatbot, be it for extraction, summarization or effective chat function with the source
  • Development of an agentic RAG system, the bot acts autonomously in several steps: understanding, researching, answering

 

Added value / goal

  • Improving the response accuracy of language models by accessing external, up-to-date and domain-specific information.
  • Answers are based on current data - even if the LLM has not been trained on it
  • Integration of in-house knowledge
  • No time-consuming training of the LLM required - context is provided at runtime
  • Sources can be provided, which increases transparency and trust
  • Can be easily integrated into existing systems (e.g. chatbots, search functions).
  • Development of intelligent systems that can autonomously solve complex tasks by combining LLMs with dynamic tools
  • More complex chatbots are ideal for use cases such as market analysis, due diligence, technical support or complex customer inquiries

AI Scraper - intelligently extract and analyze web data

Initial situation

  • A lot of important information is scattered across the web - unstructured, dynamic and in changing formats, such as information from competitors, job portals, industry directories or product pages

 

Description of the

  • Evaluation of the specific information the customer needs from the web
  • A pipeline consisting of web scraping and RAG makes it possible to automatically extract website content, index it in a meaningful way and evaluate it intelligently using LLM
  • The system uses an agent that can carry out additional research or navigate deeper into websites if required

 

Added value / goal

  • Automated competition and market monitoring
  • Flexible extraction and interpretation from structured and unstructured web content
  • RAG allows semantic search via collected data incl. source linking
  • Ideal for research, business intelligence, lead generation and much more
  • Faster decision-making through aggregated answers to natural questions

Detection and classification of damage to residential properties

​Initial situation ​

  • Damage to properties is often recorded in text form - e.g. in tenant emails, expert reports, handover reports or forms
  • Manual analysis is time-consuming and inconsistent, especially with large volumes
  • Consultation with the tenant delays the whole process and causes annoyance

 

Description ​

  • The types of damage are evaluated with the customer
  • The damage can be verified more precisely by LLM using photos in combination with free text
  • The model is developed on the basis of specific information provided by the customer as to which information must be passed on to third parties
  • Using Power Automate, the information generated can automatically create an order that can be sent to the supplier
  • Even without media discontinuity, the end-to-end process can be ensured with direct automated dispatch to suppliers, should this be desired

 

Added value / goal

  • Automated pre-processing of incoming claims texts
  • Uniform classification despite different ways of expression of the authors
  • Acceleration of claims recording and processing
  • Can be combined with image analysis for holistic assessment
  • Ideal for property management companies, insurance companies, platform providers or claims services

Object recognition on images

Initial situation ​

  • Classic object recognition is based on specialized CV models such as YOLO or Mask R-CNN
  • However, the combination of object recognition and contextual description is difficult to integrate with traditional methods
  • No automatic evaluations can be made with the images

 

Description of the

  • The type of images to be analyzed and the information that will help the customer must be defined with the customer
  • Vision-enabled LLMs can analyze images and describe them in natural language at the same time
  • They recognize and describe concrete objects - including context, location, relationships and relevance
  • Depending on the customer's specifications and requirements, the information can be processed automatically or a proposal for the next process step can be prepared

 

Added value / goal

  • Simplified object recognition + text processing in one model
  • Ideal for smart assistance systems, chatbots with image understanding, interactive reporting
  • Reduction of development effort (no separate CV architecture required)
  • High flexibility in tasks such as object description, scene analysis, quality inspection
  • Strong results even with unstructured or complex images

Recognize sign language with deep learning

Initial situation

  • People with hearing impairments use sign language to communicate - however, there are barriers in many everyday situations as not everyone can speak sign language

 

Description of the

  • Deep learning (especially CNNs and RNNs / Transformers) is used to process video input in real time to recognize hand and finger movements, facial expressions and body posture
  • The movements are translated into gestures and then transcribed into written text
  • Models are trained on annotated gesture videos, including skeleton tracking and temporal sequence processing

 

Added value / goal

  • Real-time translation of sign language into text
  • Real-time translation into text can build bridges without great technical effort
  • Enables barrier-free communication, e.g. in public authorities, stores or at the doctor's
  • Can be combined with speech synthesis for full dialog capability

Portfolio management

Initial situation

  • Complex decisions with strong dependencies are a prerequisite for portfolio management

  • Conventional systems quickly reach their limits and cannot optimize existing patterns

  • Classic models work with historical data, despite volatile markets and geopolitical events

 

Description of the

  • The RL agent learns from experience through feedback in the form of reward or punishment

  • The agent adjusts investment decisions on an ongoing basis, based on market data, volatility and risk profile

  • The training is implemented in safe simulation environments and put into production in the second step

 

Added value / goal

  • The agent learns to recognize patterns that classic models overlook

  • The risk is reduced because fewer wrong decisions are made

  • The system automatically adapts to new conditions - without manual intervention

  • Intelligent management of prices, capital and measures noticeably increases ROI