Systems / CPS / Edge AI Research

Dependable and resource-aware ML systems for networked autonomy and constrained deployment

My systems-focused work builds dependable and resource-aware ML systems for networked, resource-constrained platforms, including IoT, embedded systems, CPS, and multi-agent autonomy. I am especially interested in settings where decisions must be made under hard constraints on latency, energy, and memory, and where distribution shift and partial observability are part of normal operation rather than edge cases.

Current themes

  • Budget-aware learning and control in autonomy: Through NSF CPS AERIAL (Co-PI), I work on trustworthy and budget-aware learning-and-control for UAV swarm autonomy, along with digital-twin validation protocols and operational reliability metrics.

  • Compact graph inference for deployment: I develop low-rank and constrained training methods for graph models that make decision-time cost controllable, enabling practical tradeoffs among accuracy, memory, and latency.

  • Efficient inference for large models: I study inference-time methods such as confidence-guided stopping to reduce compute cost while maintaining task quality, particularly in resource-constrained or budget-limited settings.

  • Protocol-driven evaluation and reliability diagnostics: A recurring goal is to make reliability measurable in operational units such as errors, violations, and energy per decision using replayable traces, stress tests, and deployable evaluation pipelines.

Representative applications and settings

  • UAV autonomy and search-and-rescue settings
  • Wireless and networked decision systems
  • Resource-constrained edge and embedded environments

Systems perspective

I aim to build ML systems whose accuracy, reliability, and cost can be evaluated jointly, using protocols that remain meaningful under real operational constraints rather than idealized offline assumptions.