VenkataDurga Sai Sarat Chandra Nallabati@Sarat972007
1
Events attended
1
Submissions made
India
1 year of experience
About me
Venkata Durga Sai Sarat Chandra Nallabati Specialization: Scientific Machine Learning (SciML) & Autonomous Robotics A Robotics & Artificial Intelligence Engineering Student focused on the intersection of hardware acceleration and complex system modeling. With a research background published in IEEE Xplore, I specialize in developing high-performance autonomous frameworks and optimizing Scientific Machine Learning (SciML) models—including POD, DMD, and SINDy architectures. My expertise lies in bridging the gap between high-level AI abstractions and low-level hardware execution, specifically leveraging C++ and Python to build real-time, resilient systems for edge and cloud environments. Currently exploring the optimization of diffusion dynamics for large-scale generative tasks.
🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
Evolution Edge: A self Learning
Evolution Edge: The Self-Evolving Neural Bridge In the modern AI era, users often face a difficult trade-off between the high-performance capabilities of massive cloud models and the privacy-centric, low-latency nature of edge computing. Evolution Edge resolves this conflict by introducing a "Self-Evolving Neural Bridge," a hybrid architecture optimized specifically for the AMD AI hardware ecosystem. The system features a compact, quantized student model (such as TinyLlama) running locally on AMD Ryzen™ AI NPUs via ONNX Runtime and DirectML. This local agent handles routine tasks with sub-50ms latency, ensuring total data privacy as sensitive information never leaves the device. However, unlike traditional static edge models, Evolution Edge utilizes a Symbolic Confidence Router. This neuro-symbolic component monitors inference confidence in real-time. When a query exceeds the local model’s expertise, it triggers a secure, anonymized escalation to a "Teacher" model. The Teacher, a full Llama-3-8B instance hosted on AMD Instinct™ MI300X accelerators, processes the query using ROCm™-optimized pipelines. Beyond providing a simple answer, it performs on-the-fly Knowledge Distillation, generating high-fidelity responses and synthetic training examples. These are delivered as a "Knowledge Packet" back to the local device, where the student model integrates the new information through few-shot injection or lightweight LoRA updates. This creates a model that grows smarter with every interaction. Over time, the local model achieves progressive autonomy, reducing cloud reliance while maintaining a tiny footprint. Evolution Edge represents a paradigm shift toward lifelong learning on the edge, fully leveraging AMD’s end-to-end silicon stack from Instinct cloud power to Ryzen AI performance.
10 May 2026
