Embedded Files
LLM based Multi-agent Decision Support Framework
LLM based Multi-agent Decision Support Framework
Developing a multi-criterion-based multi-agent decision support system specifically tailored for innovation management . This system aims to integrate a diverse array of decision-making tools and methodologies, which can be utilized both as prompting techniques and as practical tools, depending on the specific use cases encountered. By leveraging the capabilities of large language models (LLMs) within a multi-agent framework, the system is designed to facilitate comprehensive and nuanced decision-making processes.
Soziomimic and Virtual User Lab
Soziomimic and Virtual User Lab
VUL is an LLM-powered user simulation platform designed to address the slow and costly nature of traditional user research, which often excludes marginalized and disability groups. By creating LLM-powered agents with customizable templates, tasks, and background data, VUL facilitates ideation, participatory innovation, market research, and strategy validation through interactive conversations. Ongoing efforts focus on enhancing agent realism through data enrichment, segmentation, and validation, as well as advancing multi-agent reasoning with debate and adaptive memory for long-term learning of customer behavior. The platform will also offers services for companies to train custom agents or use pre-trained ones. I am responsible for creating the backend agentic architecture design and development.ย
CARE-LLM: Interactive Dashboard for Enhanced Patient Interaction with Complex Medical Data
CARE-LLM: Interactive Dashboard for Enhanced Patient Interaction with Complex Medical Data
ย As a research project, we worked on project CARE-LLM which utilizes a multi-agent language model (LLM) framework to simplify complex medical information for patients with kidney diseases. By transforming intricate details into easily understandable content, the project aims to address the challenge of limited time that healthcare professionals can spend with patients, especially following hospital discharge. This initiative is carried out in partnership with the University Hospital of Cologne, Fraunhofer FIT, and Fraunhofer IAO.
Automatic Offer Generation
Automatic Offer Generation
A project aimed at automating the generation of tender documents for electronic e-commerence companies. Developed a solution for efficient product retrieval from extensive datasets based on descriptions in tender documents. Designed, developed, and deployed a semantic search system utilizing fine-tuned large language models, integrating it seamlessly into the organization's infrastructure to improve accuracy and efficiency in product retrieval.
Proactive Car HMIย
Proactive Car HMIย
As part of the technology development and research team, we have developed a prototype for cutting-edge Human-Machine Interface (HMI) technology, I handled the technical implementation. Our work led to the development of five key features as follows:ย
๐ฆ๐บ๐ฎ๐ฟ๐ ๐ง๐ฟ๐ฎ๐๐ฒ๐น ๐ฃ๐น๐ฎ๐ป๐ป๐ถ๐ป๐ด: Personalized checklists for trips, ensuring you pack all essentials.
๐ง๐ฎ๐ถ๐น๐ผ๐ฟ๐ฒ๐ฑ ๐ก๐ฎ๐๐๐ฟ๐ฒ ๐๐๐ฐ๐ฎ๐ฝ๐ฒ๐: Craving a breath of fresh air? The AI evaluates your preferences and suggests nearby nature spots that match your mood, whether itโs a secluded trail for an adventure with your furry friend or a scenic spot for a romantic picnic.
๐๐บ๐บ๐ฒ๐ฟ๐๐ถ๐๐ฒ ๐๐๐ฑ๐ถ๐ผ๐ฏ๐ผ๐ผ๐ธ ๐๐ ๐ฝ๐ฒ๐ฟ๐ถ๐ฒ๐ป๐ฐ๐ฒ: Dive into your favorite audiobook or podcast, and watch as your car generates stunning visuals based on the content, bringing stories and information to life like never before.
๐๐๐ป๐ฎ๐บ๐ถ๐ฐ ๐๐บ๐ฏ๐ถ๐ฎ๐ป๐ฐ๐ฒ: Wearing something vibrant like a red jacket? The carโs ambient lighting and dashboard accents will adjust to complement your outfit, creating a cohesive and invigorating atmosphere.
The goal of the project is to make every drive a more intelligent, intuitive, and enjoyable experience.ย
MTadv_MasterThesis_Bhupender.pdfAssessment of Mean Teacher and Prominent Adversarial Unlabeled Data for Language Classification (Master thesis)
Assessment of Mean Teacher and Prominent Adversarial Unlabeled Data for Language Classification (Master thesis)
In the realm of natural language processing, language models have displayed cutting-edge capabilities across diverse tasks. However, their vulnerability to subtle adversarial attacks, where inconspicuous text alterations disrupt performance, has been well-documented. While generating potent adversarial attacks garners more attention, research on defensive mechanisms lags behind. This master's thesis introduces a novel semi-supervised fine-tuning approach that aims to enhance a language model's robustness without compromising its original accuracy.
An empirical experiment contrasts the performance of models fine-tuned via conventional and proposed methods. This endeavor not only sheds light on the potential enhancements in language models but also contributes to uncovering the scope of defensive strategies. BERT and DistilBERT models were employed in the experiment, using two distinct datasets. Notably, the models refined through the proposed approach exhibited a 0-2% improvement in original accuracy and a significant 20-30% increase in accuracy when facing attacks compared to the conventional approach. This study thus underscores the viability of bolstering language models against adversarial challenges.
Fake_News_detection (3).pdfFake Article Detection With Weakly Supervised Learning
Fake Article Detection With Weakly Supervised Learning
Research Projectย
Research Projectย
In today's digital landscape, the proliferation of fake news is rampant, largely propelled by the emergence of social media platforms and internet-connected devices. News not only influences readers' opinions but also shapes societies, leading to a cycle of uncertainty between distinguishing authenticity from falsehood. Addressing this challenge requires supervising techniques that consistently deliver accurate information to users, a task made difficult by the scarcity of labeled training examples in real-world scenarios.
In this research paper, we confront the hurdles of combating fake news. We introduce innovative weakly supervised learning models, leveraging mean teacher, virtual adversarial training, and pseudo labelling techniques. These models incorporate three distinct noise generation methods, combined with embedding layer perturbations, to discern the legitimacy of news articles. A comprehensive comparison of these models is presented, paving the way for a new trajectory of future research.
Augmented Reality based mobile application to communicate with SIEMENS motor management device
Augmented Reality based mobile application to communicate with SIEMENS motor management device
Played a pivotal role in the development of an augmented reality-based application designed for seamless interaction with SIEMENS motor management and control devices. My primary responsibility centered on creating a robust communication library tailored for interfacing with SIMOCODE devices, leveraging the PROFINET communication protocol.
Within this scope, I crafted a communication framework that facilitated effective and reliable data exchange between the app and the SIMOCODE devices. By harnessing the power of PROFINET communication, I ensured seamless connectivity, enabling users to effortlessly engage with and control the motor management and control functionalities through the augmented reality interface.
This accomplishment represents a significant stride in enhancing user experiences within the industrial landscape, bridging the physical and digital realms through cutting-edge technology. The developed communication library underscores my expertise in creating efficient and effective connections between software applications and industrial devices, contributing to a more streamlined and interactive ecosystem for motor management and control.
Dynamic Virtualization of Industrial Device (Patented)
Dynamic Virtualization of Industrial Device (Patented)
Designed and developed a novel method and system to create and manage virtual industrial devices within an industrial network. The approach involved identifying communication patterns between industrial devices and engineering stations. This acquired knowledge was then harnessed to establish communication with the engineering station for testing software during its development phase.
This innovative prototype presented a significant opportunity: the potential to eliminate the hardware dependencies at development centers. This achievement brought about cost savings by negating the need to transport hardware from Germany to India. Furthermore, our system enabled the testing of numerous challenging scenarios that are often difficult or nearly impossible to replicate in a traditional testing environment, all made possible through virtual devices.
It holds the promise of elevating product quality, promoting a "shift left" testing strategy by addressing issues earlier in the development cycle and fostering increased feasibility for automated testing.ย
Pluggable Magnetic Connectorย
Pluggable Magnetic Connectorย
Efforts were undertaken to expedite the testing process of industrial motor management devices. To achieve this, I conceived and brought to life prototypes of magnetic connectors. These connectors were strategically designed and developed to swiftly establish connections between the inputs and outputs of the devices, facilitating efficient testing procedures.
By introducing these magnetic connector prototypes, we aimed to significantly reduce the time required for testing industrial motor management devices. The innovative design allowed for rapid and secure linking of the device components, streamlining the testing workflow. This endeavor reflects our commitment to optimizing testing practices in the industrial realm, enhancing efficiency, and contributing to the seamless operation of motor management systems.
The successful implementation of these magnetic connectors signifies our proficiency in devising practical solutions that address critical testing challenges within industrial contexts. Through this innovation, we aimed to elevate the efficiency and accuracy of device testing, ultimately enhancing the performance and reliability of industrial motor management systems.
Page updated
Google Sites
Report abuse