The rapid growth of Deep Neural Networks (DNNs) has shifted the computational focus from centralized cloud data centers to the network edge, driven by the need for low-latency inferencing, data privacy and bandwidth conservation. Contemporary edge accelerators, such as the NVIDIA Jetson series, offer specialized hardware (GPUs and Tensor Cores) that provide workstation-like performance within a small power envelope, making them ideal for local AI tasks. However, effectively leveraging these resource-constrained devices requires next-generation platforms that can manage unique hardware characteristics, such as shared CPU-GPU memory and diverse power modes, while scaling to meet the demands of Large Language Models (LLMs) and distributed environments.

The future vision for these systems involves the seamless integration of Generative AI and LLMs onto edge accelerators, addressing challenges in memory footprint and energy usage for local hosting and federated fine-tuning of LLMs. Research is moving toward Reinforcement Learning-based optimization for dynamic workload tuning and power-aware training strategies, and leveraging heterogeneous accelerators on-board Jetsons (e.g., DLA, Tensor cores) for concurrent inferencing workloads. Extending these efforts to create a unified, hardware-agnostic AI execution layer across the edge-cloud continuum is also a goal.

Students and Staff

Key Research Papers

Enterprise applications are increasingly moving toward serverless computing and Function-as-a-Service (FaaS) abstractions to achieve effortless scaling and cost efficiency. However, existing FaaS platforms are proprietary and siloed, leading to vendor lock-in and performance overheads like cold-starts and message indirection. Simultaneously, the emergence of Hybrid Quantum-Classical (HQC) workloads necessitates middleware that can orchestrate tasks across traditional classical servers and noisy quantum backends.

The vision for ongoing research focuses agentic frameworks that can orchestrate long-running agentic tasks and advanced memory summarization to prevent context window explosion. We are also exploring serverless within the edge-cloud continuum to support dynamic AI workloads that can adaptively migrate. In quantum platforms, the goal is to design intuitive abstractions to effectively design quantum-cloud and quantum-HPC applications, and middleware to automate the hardware optimizations that navigate trade-offs between cost, execution time and fidelity as we move toward quantum advantage.

Students and Staff

Key Research Papers

Massive graphs representing social networks, financial transactions and road networks are inherently dynamic and often reach billion-scale entities, exceeding the capacity of single-machine memory. These temporal graphs assign lifespans to vertices and edges, requiring analytics that can navigate structure and time concurrently without redundant recomputations of historical snapshots. Currently, there is a lack of scalable abstractions that can handle both time-independent and time-dependent algorithms on these evolving structures.

The future research vision includes extending these models to out-of-core disk-based GNN training and inferencing to scale beyond distributed memory, and temporal GNN inferencing with topology changes. We are also examining scalable mining of graph patterns in a streaming context.

Students and Staff

Key Research Papers

Distributed systems research is vital for solving high-impact problems in real-world environments, ranging from urban smart city infrastructures to autonomous drone fleets. These domains require low-latency, resilient data storage and scheduling heuristics that can adapt to high-rate data streams and mobile edge resources. For instance, tracking objects across a network of urban cameras or managing financial transactions at a national scale involves handling millions of mutations per second while maintaining deterministic outcomes.

Students and Staff

Key Research Papers