CRAWO framework optimizes AI pipelines across heterogeneous edge devices

CRAWO tackles the challenge of deploying AI pipelines across diverse edge devices—from low-power microcontrollers to GPU-equipped nodes—without the inefficiencies of current orchestration platforms. Developed by a team of researchers from Brazilian universities, the framework separates allocation intelligence from execution using a control-loop model. It manages placement decisions, state, and inter-stage data flows while instantiating services on edge nodes. The hardware-aware allocator includes a pluggable multi-criteria decision layer that uses real-time infrastructure metrics (e.g., CPU, memory, network) to optimize workload placement dynamically.

Built on a microservices architecture running on K3s (lightweight Kubernetes), CRAWO uses Custom Resource Definitions (CRDs) for domain modeling and a dedicated Kubernetes operator for state reconciliation. In a vehicle surveillance test case involving license plate recognition, the system demonstrated better workload distribution across edge nodes and significantly reduced dependency on centralized cloud processing. This is critical for latency-sensitive smart city applications like traffic monitoring, public safety, and real-time analytics where milliseconds matter.