Sashi Kiran Kaata

Sashi Kiran Kaata

Auto-generated profile, based on publicly available information

Sashi Kiran Kaata

Independent Researcher,Glendale,AZ,USA,85305

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
3 September 2025
Verified
Conference Paper
Closed access
OKOhm Hareesh Kundurthy
Ohm Hareesh Kundurthy
SKSashi Kiran Kaata
Sashi Kiran Kaata
SVSrinivas Vikram
Srinivas Vikram
RSRamesh Somayajula
Ramesh Somayajula
RGRajesh Gangavarapu
Rajesh Gangavarapu
The emergence of the recent superpowers of Generative AI (GenAl) has transformed content generation and decision making as well as autonomous processes. Nevertheless, it seems to be difficult to use large-scale foundation models in distributed and resource-constrained settings, i.e., in edge devices, as the scalability, communication latency, and data confidentiality become challenging. In this paper, a groundbreaking model called MicroLLMs, which provides Lightweight Generative AI, will traverse to enable effective, safe, and scalable intelligence on the cloud and edge. The framework implements modular model compression, federated distillation, and zero-trust data pipelines to enable on-device inference at asset-level performance and privacy preservation. By combining the transformation miniaturization with privacy-preserving analytics, the suggested architecture can achieve sub-second latency and minimal computational overhead with generative fidelity across heterogeneous environments. Also, adaptive orchestration mechanisms synchronize models, tolerance fault, and have policy-based control in hybrid deployments. Empirical testing over production-level edge applications, such as smart surveillance and industrial IoT, reveals that inferencing is 47 percent faster, throughput increased by 38 percent and there is substantial energy savings, compared to other traditional GenAl implementations. This framework is the foundation of democratizing GenAl capabilities to realize mobile, embedded and cloud-edge systems in a secure and long-lasting fashion.

10 December 2025 in IEEE Access
Verified
Journal Article
Open access
SJSiva Tanay Akash Jagarapu
Siva Tanay Akash Jagarapu
SKSashi Kiran Kaata
Sashi Kiran Kaata
SBShaily Bajpai
Shaily Bajpai
RMRAVIRAJA HOLLA M
RAVIRAJA HOLLA M
Accurate segmentation of cirrhotic livers from Magnetic Resonance Imaging (MRI) remains a critical challenge due to irregular liver morphology, heterogeneous tissue textures, and modality-specific artifacts. In this study, we propose DualBranchNetwork3D, a hybrid architecture that integrates a residual convolutional encoder for local texture modeling with Swin Transformer blocks for global context learning, combined via multi-scale feature fusion. The model employs a composite loss function incorporating Dice, Focal, Tversky, and boundary losses to address class imbalance and complex anatomical variations. Evaluated on the CirrMRI600+ dataset, comprising 310 T1-weighted (T1W) and 318 T2-weighted (T2W) MRI scans, our approach achieved state-of-the-art performance with Dice scores of 0.9453 on T1W and 0.9018 on T2W, outperforming existing baselines by 6.6% and 4.3%, respectively. These results validate the effectiveness of hybrid CNN-Transformer architectures for cirrhotic liver segmentation and highlight the potential for integrating such models into clinical workflows for fibrosis staging, volumetric assessment, and treatment planning.

31 October 2025
Verified
Conference Paper
Closed access
DGDipesh Garg
Dipesh Garg
SKSashi Kiran Kaata
Sashi Kiran Kaata
DBDinesh Reddy Bhumireddy
Dinesh Reddy Bhumireddy
BBBharath Kumar Bolla
Bharath Kumar Bolla
The adoption of cloud computing has led to escalating operational costs, prompting organizations to seek cost-optimization strategies. Amazon Web Services (AWS) Spot Instances offer substantial savings of up to 90 % compared to on-demand pricing. However, they present reliability challenges owing to price volatility and the potential for instance termination. This study addresses the critical challenge of developing an optimal bidding strategy that balances cost savings and operational reliability. Using a comprehensive dataset of over 27 million AWS Spot Instance pricing records across 11 regions, multiple time-series forecasting models, including naive methods, exponential smoothing variants, Holt-Winters models, and SARIMA, were systematically evaluated. Rather than focusing solely on predictive accuracy, business-oriented evaluation metrics (uptime percentage and relative cost) were used to assess practical performance in real-world scenarios. The results demonstrate that traditional forecasting accuracy metrics fail to adequately reflect the operational requirements. Our proposed Custom 25 model, which applies a 25 % safety buffer above the Holt-Winters multiplicative forecasts, consistently achieved 100 % uptime across diverse instance types (m4.xlarge, i3.xlarge, i3.2xlarge, m4.large) while maintaining cost-effectiveness. This approach provides $\text{7 0 - 8 0 \%}$ savings compared to on-demand pricing while ensuring uninterrupted service availability. This study contributes a practical, data-driven bidding framework that enables organizations to confidently integrate Spot Instances into mission-critical workloads, effectively addressing the traditional trade-off between cost optimization and service reliability in cloud computing environments.

4 July 2025
Verified
Conference Paper
Closed access
RSRohan Shahane
Rohan Shahane
Shazia HassanShazia Hassan
Shazia Hassan
NGNandita Giri
Nandita Giri
SKSashi Kiran Kaata
Sashi Kiran Kaata
VKVijayakumar. Krishnapillai
Vijayakumar. Krishnapillai
Generating correct and flexible models of realistic, partially unknown worlds is an old goal in artificial intelligence. Developments in representation learning, generative modelling, inventions of self-supervised methods mean that AI systems have attained inner models of the world that allow them to perform nearly the same as human perception regarding real-world dynamics without direct supervision. In this paper, an innovative training scheme to train AI agents to anticipate a reality by learning self-built portrayal of their environment is covered. Using variational autoencoders and generative adversarial networks, our method allows agents to learn underlying structure and predict future states and in addition to being able to represent a large number of complex realistic scenarios. We experimentally analyse the efficacy of this approach in the domains of synthetic and physical preparations that reinforce the realization of agents outfitted with self-built world models giving better generalization, predictive precision, and utilization of resources. With references to experimental results, the models are likely to contribute to increasing real-time decision making, autonomous control as well as creation of virtual environment in a wide selection of practical applications. Lastly, we mention some of the main limitations such as representation fidelity, scalability, and computational complexity and future directions depending on which simulations could be made to use better representations and achieve results that could be deployed in dynamic real-world environments more easily.