Jan-Hendrik Witte

Glasfaser Nordwest · since 2025

Academic work · University of Oldenburg, 2018–2025

• 2024 — 2025

  DocNexus

  Volkswagen AGLead

  Project DocNexus addresses the challenges faced by Internal Audit departments in managing and utilizing vast amounts of unstructured and semi-structured data, including process standards, audit reports, and other audit-relevant documentation. The project recognizes that traditional methods of searching and compiling necessary documents for audit preparation are time-consuming and labor-intensive, particularly due to the limitations of simple keyword searches and query-based information retrieval. To solve these challenges, DocNexus leverages state-of-the-art AI technologies, including Generative AI, multimodal models, Large Language Models (LLMs), and Retrieval-Augmented Generation (RAG). The research focuses on developing innovative solutions for processing and extracting information from large volumes of unstructured data, ultimately aiming to enhance the efficiency and effectiveness of daily audit operations through intelligent information processing and retrieval systems.

  Highlights

  • Innovative application of state-of-the-art AI technologies to transform internal audit processes
  • Smart information extraction and processing of complex, unstructured audit documentation
  • Development of an intelligent system to significantly reduce document search and compilation time

  Collaborators

  • University of Oldenburg
  • Volkswagen AG

  Links

  • Website
• 2023 — 2024

  Kalkulation.KI

  Strabag AGContributor

  Project Kalkulation.KI is a collaborative research initiative between the Department of Business Information Systems/VLBA at Carl von Ossietzky University Oldenburg and STRABAG AG. The project aims to significantly automate and accelerate the bid processing and calculation for construction projects using proven artificial intelligence methods. Through the implementation of advanced text mining and machine learning techniques, the project develops an intelligent IT solution that integrates seamlessly with STRABAG's process control software.

  Highlights

  • Development of an AI-powered system for automated construction bid processing and calculation
  • Integration of advanced text mining methods to enhance bid preparation efficiency
  • Implementation of machine learning algorithms for intelligent pattern recognition and solution filtering

  Collaborators

  • University of Oldenburg
  • Strabag AG

  Links

  • Website
• 2020 — 2024

  DigiSchwein

  BMELLead

  The DigiSchwein project was established to explore the potential of digital technologies in pig farming. Focusing on current challenges such as early disease detection, tail biting, birth monitoring, and nutrient flows, the project aims to develop a digital early warning system for practical farming operations. This farm management system continuously collects real-time data about animals, environment, and facilities through various types of sensors. By networking and analyzing the sensor data, farmers receive current status updates, forecasts, and action recommendations for their operations, effectively supporting their daily work.

  Highlights

  • Deep learning models for pig detection and tracking
  • Behavior analysis algorithms
  • Real-time monitoring capabilities

  Collaborators

  • Lower Saxony Chamber of Agriculture
  • University of Oldenburg
  • University of Veterinary Medicine Hannover, Foundation
  • University of Kiel
  • OFFIS Institute
  • Thünen Institute

  Links

  • GitHub
  • Demo
  • Website
• 2018 — 2020

  TRACE

  Volkswagen AGContributor

  Project TRACE, a collaborative research initiative between VLBA and Volkswagen AG, explores the practical applications of Data Science, with a particular focus on analytical methods for connecting independent market data. The project aims to develop a data-driven perspective on the corporate environment and markets by exploring and analyzing various data sources, including procurement and financial markets. The research addresses the challenges of semantic, structural, and technical heterogeneity in data integration through advanced Data Science concepts and methodologies. By implementing sophisticated techniques such as record linkage procedures, the project works to identify and connect datasets representing the same real-world entities. The ultimate goal is to conceptualize, prototype, and evaluate a system that enables a comprehensive data-driven view of the corporate environment and relevant markets.

  Highlights

  • Development of innovative data integration methods for market analysis
  • Implementation of advanced record linkage techniques for heterogeneous data sources
  • Creation of a prototype system for comprehensive market intelligence

  Collaborators

  • University of Oldenburg
  • Volkswagen AG

  Links

  • Website
• 2026 — 2029

  RECOMPOSE

  Proposal

  Development of an AI-powered platform architecture to foster critical AI literacies among students and teachers, focusing on music education. The project creates an explainable generative AI system for music generation that enables students to understand and manipulate decision-making processes within AI models, promoting informed and reflective use of AI technologies in education.

  Contributed to the project proposal.
• 2024 — 2027

  EduAID

  Proposal

  Integration of AI-related tools and practices in education through a microcredentials approach targeting teachers, school heads, and educators. The project develops an AI-powered app and digital platform for personalized learning, establishes a framework for micro-credentials recognition, and conducts pilot programs to enhance digital literacy and teaching quality.

  Contributed to the project proposal.
• 2024 — 2027

  SHIELD

  Proposal

  A 36-month Erasmus+ project combating gender-based violence (GBV) among young people in Europe, particularly stalking, harassment, dating violence, and online grooming. The project develops an AI-powered chatbot to detect GBV indicators and creates an interactive online platform providing training modules and best practices for youth workers, educators, and caregivers.

  Contributed to the project proposal.
• 2023 — 2025

  Transparency in Pig Production (TiPP)

  Proposal

  Optimization of transparency and traceability "from farm to fork" in regional pork supply chains using digital strategies. The project employs RFID-supported sensor technology and explores Self Sovereign Identity (SSI) concepts to enhance data security and flexibility in bidirectional data exchange, ultimately deriving transparency indices for consumers.

  Contributed to the project proposal.

• 2026

  Recent Advances in Information Systems: Proceedings of the ICRAIS, Pointe aux Piments, Mauritius, 10–12 September 2025

  Marx Gómez, J.; Sungkur, R. K.; Pudaruth, S.; Witte, J.-H.; Schumann, G. (Eds.)

  Progress in IS · Springer ChamEdited Volume

  • DOI
  • Publisher
  About the book

  The proceedings of the 2025 International Conference on Recent Advances in Information Systems, held in Pointe aux Piments, Mauritius, and organised jointly by the University of Oldenburg and the University of Mauritius. Its 64 papers by authors from research and industry cover information systems, artificial intelligence and machine learning, among them explainable and sustainable AI, security and privacy, educational technology and applied computer vision.

• 2026

  Automated Detection and Tracking of Individual Piglets for Real-Time Behavioral Monitoring

  Witte, J.-H.; Ahrens, D.-N.; Traulsen, I.; Marx Gómez, J.

  International Conference on Recent Advances in Information Systems (ICRAIS 2025)Conference Paper

  • DOI
  • Publisher
  Abstract

  A camera-based system that detects and follows individual piglets in farrowing pens using standard RGB cameras. Several deep learning detectors and trackers were compared on what makes these pens hard: piglets huddle together, hide one another and move unpredictably. The best combination, a YOLOv5l detector with SORT tracking, reached 98.6% MOTA and 93.9% IDF1 at 38 frames per second under test conditions. From the tracks the system derives distance travelled, average speed and idle time for each piglet, which match manual annotations closely (mean squared error of about 0.01). Stationary and tightly clustered piglets remain the open problems; the datasets are publicly available.

• 2026

  Large Language Models for Qualitative Risk Analysis in Banking: A Systematic Review

  Witte, J.-H.; Eggers, H.

  International Conference on Recent Advances in Information Systems (ICRAIS 2025)Conference Paper

  • DOI
  • Publisher
  Abstract

  A systematic review of how large language models support qualitative risk analysis in the credit risk management of banks. The literature falls into five application areas: credit scoring and default forecasting, broader risk assessment, document analysis and report generation, fraud detection, and regulatory compliance. Across them, models from general-purpose LLMs such as GPT-4 to finance-specific ones such as FinBERT improve predictions by combining unstructured text with conventional financial figures, and make the work more efficient. Data quality and privacy, explainability, computing cost and regulation remain open. The review closes with what this means for banks and an agenda for research: explainable financial LLMs, privacy-preserving methods, leaner architectures, hybrid models and regulatory frameworks.

• 2024

  Tail Posture as a Predictor of Tail Biting in Pigs: A Camera-Based Monitoring System

  Witte, J.-H.; Heseker, P.; Probst, J.; Traulsen, I.; Kemper, N.; Marx Gómez, J.

  11th European Conference on Precision Livestock FarmingConference Paper

  • Code
  • ResearchGate
  Abstract

  This study presents a novel camera-based monitoring pipeline for analyzing pig tail posture , a crucial indicator of tail biting risk in pig livestock farming. Utilizing a multi-step approach, the system first employs a YOLOv8 model for pig detection, followed by an EfficientNetV2 model to classify pigs into 'lying' and 'not lying' postures. This classification aids in focusing on relevant pigs for tail posture analysis by filtering out ambiguous tail postures of 'lying' pigs from the monitoring process. Subsequently, another YOLOv8 model detects upright or hanging tail postures on the pig detections classified as 'not ly-ing', enhancing the accuracy of the monitoring process. The effectiveness of this pipeline is evaluated based on video recordings of seven piglet rearing batches across three different pens, including 14 pens with observed tail biting incidents and 7 control groups. Retrospectively , the system could have provided early alerts for impending tail biting outbreaks at least one day in advance in 61.5% of cases and indicated a 41.5% reduction in upright tail postures within seven days leading up to an outbreak. This study demonstrates significant promise in improving animal welfare and operational efficiency in farms. By providing a proactive tool for assessing tail biting risks, this system could reduce the reliance on traditional practices like tail docking, paving the way for more ethical and sustainable farming methods. With exception for the video data used for evaluating, all of the applied models, datasets, as well as the monitoring pipeline outputs and its implementation code are publicly available.

• 2024

  Image-based activity monitoring of pigs

  Witte, J.-H.; Marx Gómez, J.

  44. GIL - Jahrestagung, Biodiversität fördern durch digitale LandwirtschaftConference Paper

  • PDF
  • DOI
  • Code
  • ResearchGate
  Abstract

  In modern pig livestock farming, animal well-being is of paramount importance. Monitoring activity is crucial for early detection of potential health or behavioral anomalies. Traditional object tracking methods such as DeepSort often falter due to the pigs' similar appearances, frequent overlaps, and close-proximity movements, making consistent long-term tracking challenging. To address this, our study presents a novel methodology that eliminates the need for conventional tracking to capture activity on pen-level. Instead, we segment video frames into predefined sectors, where pig postures are determined using YOLOv8 for pig detection and EfficientNetV2 for posture classification. Activity levels are then assessed by comparing sector counts between consecutive frames. Preliminary results indicate discernible variations in pig activity throughout the day, highlighting the efficacy of our method in capturing activity patterns. While promising, this approach remains a proof of concept, and its practical implications for real-world agricultural settings warrant further investigation.

• 2024

  Image-based Tail Posture Monitoring of Pigs

  Witte, J.-H.; Heseker, P.; Probst, J.; Traulsen, I.; Kemper, N.; Marx Gómez, J.

  57th Hawaii International Conference on System SciencesConference Paper

  • PDF
  • DOI
  • Code
  • ResearchGate
  Abstract

  Tail biting presents a significant challenge in conventional pig farming, impacting animal welfare and farmers' economic viability. This paper introduces a novel approach for image-based tail posture monitoring, a potential early indicator of tail biting outbreaks. Our two-step tail posture detection approach, consisting of an initial pig detection and a subsequent tail posture detection step, shows significant improvements compared to previous methods. To mitigate ambiguity, our pipeline incorporates an EfficientNetV2 image classification model, filtering out lying pigs in the tail posture monitoring process. When applied to video sequences containing tail biting incidents, our method effectively captures the shift in tail posture from predominantly upright to hanging preceding outbreaks. Our findings offer a promising foundation for an early warning system to aid undocked pig husbandry, improve animal welfare, and provide targeted insights for farmers. The proposed approach demonstrates the potential for real-world applications, fostering proactive interventions to mitigate tail biting.

• 2022

  Introducing a New Car-Sharing Concept to Build Driving Communities for Work-Commuting

  Witte, J.-H.

  In: Digital Transformation for SustainabilityBook Chapter

  • DOI
  • Publisher
  • ResearchGate
  Abstract

  South Africa is currently facing various mobility problems. On the one hand, more and more people are moving up into the middle class, which increases the number of private vehicles on the roads and leads to congestion, increased pollutant emissions, and overcrowded cities. On the other hand, public transportation in the form of buses and trains are hardly used due to their current state in terms of safety, efficiency, or availability, making low-cost trips much more difficult for people without private vehicles. The daily commute to work becomes particularly challenging due to crowded roads and the lack of public transportation alternatives. During the 2018 HEdIS Summer School, an international group of students from South & West Africa and Germany collaborated to develop innovative ideas to address these issues. Using the design thinking process, a new car-sharing concept was developed to enable the potential formation of consistent driving communities. The goal is to match users who live in close proximity to each other and are employed by the same company or by companies that are near each other. Safety is ensured by only giving access to users who are currently employed by a company with which the platform currently has a partnership. By effectively matching people with the same work commute, it is intended to provide an alternative commuting option that is safer, more efficient, and less time-consuming compared to other available modes of transportation and to overall reduce the number of vehicles on the road to address traffic congestion.

• 2022

  Structured and Targeted Communication as an Enabler for Sustainable Data Science Projects

  Kruse, F.; Kessler, R.; Witte, J.-H.

  In: Digital Transformation for SustainabilityBook Chapter

  • DOI
  • Publisher
  • ResearchGate
  Abstract

  In the modern, highly digitized world, the use of data plays an increasingly important role. In order to extract information from data and create value, data products must be created and developed. In the field of sustainability, for example, in the smart city context, it can also be observed that Data Science is becoming more and more central. There is a need for data scientists and stakeholders to actively collaborate to create valuable data products based primarily on the data. In general, it is the job of data scientists to develop the data product and the job of decision-makers to use the data product to support decisions. Therefore, effective communication between data scientists and stakeholders is a crucial success factor in data science projects. Communication problems can occur at the interface between data scientists and stakeholders. To date, there is no approach to implement and facilitate efficient communication processes between data scientists and stakeholders. This chapter presents an approach for efficient communication processes between stakeholders and data scientists called the Data Product Profile. The concept was developed based on existing literature and expert interviews and evaluated with practitioners.

• 2022

  Using Deep Learning for automated birth detection during farrowing

  Witte, J.-H.; Gerberding, J.; Lensches, C.; Traulsen, I.

  EnviroInfo 2022Conference Paper

  • PDF
  • Publisher
  • ResearchGate
  Abstract

  Pig livestock farming has been undergoing major structural change for years. The number of animals per farm is constantly increasing, while competition is becoming more intense due to volatile slaughter prices. Sustainable, welfare-oriented livestock farming becomes increasingly difficult under these conditions. Studies have shown that animal-specific birth monitoring of sows can significantly reduce piglet losses. However, continuous monitoring by human staff is inconceivable, which is why systems need to be created that assist farmers in these tasks. For this reason, this paper aims to introduce the first step towards an automated birth monitoring system. The goal is to use deep learning methods from the field of computer vision to enable the detection of individual piglet births based on image data. This information can be used to develop systems that detect the beginning of a birth process, measure the duration of piglet births, and determine the time intervals between piglet births.

• 2022

  Using Deep Learning for Automated Tail Posture Detection of Pigs

  Witte, J.-H.; Gerberding, J.; Marx Gómez, J.

  11th International Conference on Data AnalyticsConference Paper

  • PDF
  • Publisher
  • ResearchGate
  Abstract

  Tail biting is one of the biggest problems in pig livestock farming. One indicator that can be observed before an outbreak is the change in tail posture. Studies have shown that days before a tail biting outbreak, a steady increase in hanging tail postures can be observed. A continuous monitoring of this indicator could therefore be used to inform farmers of potential problems arising within respective pens. This paper therefore presents a first step in the development of automated monitoring systems for early detection of tail biting indicators by evaluating different approaches for tail posture detection using image data and Deep Learning. Using a dataset consisting of 1000 annotated images, different YOLOv5 object detection models were trained to detect upright and hanging tail postures. The results show that there are significant differences in performance for the detection of upright and hanging class. To further investigate the problem, an EfficientNetv2 image classification model was trained to examine if similar performance differences for the two classes could be observed. Considered in isolation, these differences could be mitigated. However, potentials could not be utilized, as the results of the comparison of the one-step detection of tail posture using YOLOv5 and the introduced two-step detection using YOLOv5 for tail detection and EfficientNetv2 for tail posture classification shows. Based on the discussion of the possible explanations for the inferior performance as well as the summary of the key findings of this paper, we present approaches that can be used as a basis for future research.

• 2022

  Introducing a New Workflow for Pig Posture Classification Based on a Combination of YOLO and EfficientNet

  Witte, J.-H.; Marx Gómez, J.

  55th Hawaii International Conference on System SciencesConference Paper

  • PDF
  • DOI
  • Code
  • ResearchGate
  Abstract

  This paper introduces a pipeline for image-based pig posture classification by applying YOLOv5 for pig detection and EfficientNet for subsequent pig posture classification into 'lying' and 'notLying'. A high-quality dataset consisting of 5311 heterogeneous images from different sources with 78215 bounding box annotations was created. The bounding box annotations were then used to create a separate dataset for image classification, consisting of 9209 and 7855 images for each 'lying' and 'notLying'. The YOLOv5 model achieves an AP of 0.994 for pig detection, while EfficientNet achieves a precision of 0.93 for pig posture classification. Comparing the results of the proposed method with other approaches found in literature, it shows that significant improvements in terms of accuracy can be achieved by splitting the classification of pig posture into separate models. This research provides a foundation for the continued development of real-time monitoring and assistance systems in pig Precision Livestock Farming.

• 2021

  Evaluation of Deep Learning Instance Segmentation Models for Pig Precision Livestock Farming

  Witte, J.-H.; Gerberding, J.; Melching, C.; Marx Gómez, J.

  24th International Conference on Business Information SystemsConference PaperBest Paper Award

  • PDF
  • DOI
  • ResearchGate
  Abstract

  In this paper, the deep learning instance segmentation architectures DetectoRS, SOLOv2, DETR and Mask R-CNN were applied to data from the field of Pig Precision Livestock Farming to investigate whether these models can address the specific challenges of this domain. For this purpose, we created a custom dataset consisting of 731 images with high heterogeneity and high-quality segmentation masks. For evaluation, the standard metric for benchmarking instance segmentation models in computer vision, the mean average precision, was used. The results show that all tested models can be applied to the considered domain in terms of prediction accuracy. With a mAP of 0.848, DetectoRS achieves the best results on the test set, but is also the largest model with the greatest hardware requirements. It turns out that increasing model complexity and size does not have a large impact on prediction accuracy for instance segmentation of pigs. DETR, SOLOv2, and Mask R-CNN achieve similar results to DetectoRS with a parameter count almost three times smaller. Visual evaluation of predictions shows quality differences in terms of accuracy of segmentation masks. DetectoRS generates the best masks overall, while DETR has advantages in correctly segmenting the tail region. However, it can be observed that each of the tested models has problems in assigning segmentation masks correctly once a pig is overlapped. The results demonstrate the potential of deep learning instance segmentation models in Pig Precision Livestock Farming and lay the foundation for future research in this area.

Four courses, taught in German and English, and eight theses supervised, five Master's and three Bachelor's, between 2021 and 2025.