Bhanuprakash Madupati

Mr. Bhanuprakash Madupati

orcid

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
4 July 2025
Verified
Conference Paper
Closed access
AJAnil Kumar Jonnalagadda
Anil Kumar Jonnalagadda
BMBhanuprakash Madupati
Bhanuprakash Madupati
RVRohith Varma Vegesna
Rohith Varma Vegesna
SVSantosh Kumar Vududala
Santosh Kumar Vududala
LLMs have brought new, amazing abilities for understanding language, generating it and making decisions. Yet, there are serious concerns about data privacy, the ability to scale LLMs and how different components of a cloud-native system interact. The paper outlines a new Federated Data Modelling (FDM) framework specifically for making use of LLMs in secure and efficient distributed cloud settings. The framework achieves decentralized training, prevents data being leaked and meets the requirements of data residency laws by using federated learning and dynamic schema harmonization with container orchestration. Moreover, the proposed FDM technique relies on zero-trust security, confidential computing and Kubernetes-native operations to provide isolation, watching and traceability among the various tenants. On typical benchmark datasets, the approach shown here performs better in terms of privacy, how quickly the model learns and how quickly it may be used in practice compared to centralized training. By using this study, AI service providers can ensure their LLM service is trustworthy and safe for IAP use in healthcare, finance and government.

4 July 2025
Verified
Conference Paper
Closed access
BMBhanuprakash Madupati
Bhanuprakash Madupati
AJAnil Kumar Jonnalagadda
Anil Kumar Jonnalagadda
SVSantosh Kumar Vududala
Santosh Kumar Vududala
RVRohith Varma Vegesna
Rohith Varma Vegesna
Advances in LLMs have altered AI analytics, yet they are not yet easily brought into scalable, secure and efficient engineering systems—especially in settings with various cloud platforms. The paper offers a new design for LLM-driven data engineering processes that resolves major issues in collecting, processing and publishing data on the cloud. A proposed framework includes federated orchestration, encryption of all data and access control policies without trust, to achieve proper data privacy and adherence to rules, as well as ensure fast realtime processing of both unstructured and semi-structured data. In addition, we introduce a new scheduler and metadata manager to automatically optimize things like building prompts, training and inferencing using machine learning models. Extensive study using common benchmark sets on AWS, Azure and GCP demonstrates that there are 35% fewer pipeline problems, cost is reduced by 28% and the system performs more effectively and is more resilient. This research outlines how secure and scalable use of AI systems powered by LLMs can be spread across various multi-cloud platforms.

1 March 2025
Verified
Journal Article
Open access
PDF file
BMBhanuprakash Madupati
Bhanuprakash Madupati
AJAnil Kumar Jonnalagadda
Anil Kumar Jonnalagadda
SMSupriya Madupathi
Supriya Madupathi
Massive advancements in NLP have been noted in Large Language Models (LLMs), among which OpenAI's ChatGPT and DeepSeek V1 are the most recognized. The following is a comparison of these models' architecture, efficiency, and performance. ChatGPT's development is based on a dense transformer with an Enhanced Transformer Architecture. It has evolved through different generations, such as GPT-3.5, GPT-4, and the optimized versions, O3 and O1. DeepSeek V1 follows an architecture with a Mixture-of-Experts (MoE), where some of the model's parameters are toggled to perform computations more efficiently. Their structural characteristics, inference speed, and parameter utilization are compared, along with real-world applications. It is observed that ChatGPT demonstrates state-of-the-art performance in general NLP tasks, while DeepSeek V1 excels in specific expert tasks. This raises the issue of computational cost, which was found to be positively correlated with the model's versatility, indicating that the best choice of a model may depend on its intended use. The paper concludes with general suggestions for future improvements in LLMs, including potential directions such as hybrid models based on MoE-Transformer architecture.

25 April 2025
Verified
Conference Paper
Open access
SGShubham Gupta
Shubham Gupta
KVKusumakumari Vanteru
Kusumakumari Vanteru
SKSrinivas Reddy Kosna
Srinivas Reddy Kosna
BMBhanuprakash Madupati
Bhanuprakash Madupati

1 January 2025 in SSRN Electronic Journal
Verified
Preprint
Open access
BMBhanuprakash Madupati
Bhanuprakash Madupati
ICS (Industrial Control Systems) are the backbone of power, water, and manufacturing, amongst other critical infrastructure sectors. Like everything else, traditional ICS is evolving with the modern Information and Communication Technologies (ICT) getting integrated into its stack, exposing itself to potential cyber-attacks. Because of the real-time operational requirements and legacy technology, traditional security methods are frequently ineffective in protecting ICS. Machine Learning (ML) techniques are the major solution to improve Intrusion Detection Systems (IDS) or Intrusion Prevention Systems (IPS). This article is a survey paper in which we discussed the use of ML for cybersecurity, such as anomaly detection, integrations with crypto, and adversarial attacks within ICS. Through a discussion of the challenges and future directions around the deployment of ML models inside the ICS environment, this work aims to demonstrate how advanced technologies can help safeguard critical infrastructure against adaptive threats. More focus is given to enhancing the robustness and scalability of the ML models across device/ICS networks.

9 July 2025
Verified
Conference Paper
Closed access
BMBhanuprakash Madupati
Bhanuprakash Madupati
KAKaran Alang
Karan Alang
SKSrikanth Kamatala
Srikanth Kamatala
VPVenkata Reddy Pasam
Venkata Reddy Pasam
AJAnil Kumar Jonnalagadda
Anil Kumar Jonnalagadda
Modern data lake houses encounter substantial challenges due to their rapid increase in heterogeneous and unstructured data when it comes to metadata administration and classification and semantic discovery. Such massive and diverse datasets exceed the limits of traditional schema-based methods as well as rigid rule-based systems. This research develops an AI-based method for intelligent classification and automated metadata extraction through Large Language Models. The system uses pre-trained LLMs' contextual understanding and transfer learning abilities to accurately extract metadata from documents logs and multimedia items which have semi-structured and unstructured formats. The framework incorporates a zero-shot learning method from LLMs which teams up with supervised fine-tuning for domainspecific tagging. The validated architecture shows better performance and adaptivity through achieving superior accuracy than traditional NLP-based metadata engines when applied to real-world lakehouse dataset storage. Experimental findings demonstrate that LLMs create substantial enhancements in metadata development as well as search functions and data protection metrics in modern big data systems. This proves LLMs can bring substantial changes to future big data platforms.