Moustafa’s research investigates autonomous orchestration of complex systems under uncertainty, i.e., how to build systems that coordinate many interdependent components, adapt when conditions change, and produce reliable outcomes without constant human intervention. This question has guided his work across four domains: high-performance computing, edge-native ML systems, mission-critical autonomous aviation, and compound AI systems for machine learning engineering. Three themes persist: the limiting factor for autonomy is context, the system’s ability to understand its environment, track provenance, and reason about dependencies it cannot directly observe; autonomous orchestration at scale requires explicit constraint specification and runtime verification, not just capable components; and systems architecture is a multiplier, i.e., the right orchestration framework makes components dramatically more effective than they are alone.

High Performance Computing & Distributed Systems

Moustafa’s doctoral research addressed the first instance of this orchestration problem: enabling computational scientists to execute complex workflows across heterogeneous infrastructure, including supercomputers, clouds, grids, and clusters, where manual resource orchestration is infeasible. He developed a constraint-based, declarative programming model and an autonomic composition framework that dynamically federates distributed resources based on application behavior, resource availability, and quality-of-service objectives. This work enabled computational scientists to execute complex workflows across 100,000+ processors without manual orchestration. The resulting CometCloud framework supported applications ranging from cancer detection at Rutgers Cancer Institute (reducing analysis time from months to minutes), wildfire prediction at UCSD, oil reservoir simulation at UT Austin, molecular dynamics, dissipative particle dynamics, and metagenomics workflows. This work was recognized with nomination for the ACM Doctoral Dissertation Award and awards at IEEE SCALE 2011 and ACM UCC 2015. Multi-cloud container orchestration work with IBM Research was featured in Fortune Magazine.

Edge Computing & ML Systems

At UC Berkeley’s RISELab, the orchestration question shifted from infrastructure to ML systems operating at the network edge. Moustafa’s postdoctoral research examined the design of ML-driven IoT systems that must operate continuously despite network failures, resource constraints, and decentralized trust assumptions. He led development of the XBOS-DR platform, a distributed operating system for smart buildings deployed across hundreds of commercial buildings in California for real-time monitoring and energy management. His work on serverless computing across the cloud–edge continuum provided the deployment infrastructure for these systems, addressing how ML models are placed and executed across distributed cloud and edge nodes. On top of this infrastructure, he developed ML models to simulate building thermal response (digital twins) and investigated deep reinforcement learning for coordinating HVAC consumption and smart vehicle charging in response to grid demand-response events. These systems operated under partially observable conditions, e.g., building occupancy, external weather, grid state, making real-time decisions while respecting physical and safety constraints. The platform produced a public research dataset (mortardata.org). He also contributed to WAVE, a decentralized authorization framework enabling secure, scalable IoT operation without centralized trust, published at USENIX Security 2019 and demonstrated on hundreds of deployed devices. This research included a widely cited comprehensive tutorial on blockchain systems published in ACM Computing Surveys.

Mission-Critical AI Infrastructure

At NASA Ames Research Center, the same coordination challenge extended into domains where failures carry safety-critical consequences. As a Technical Director, Moustafa led research and engineering on infrastructure for autonomous systems requiring real-time coordination under strict security, latency, and reliability constraints.

He designed and led development of the Data & Reasoning Fabric (DRF), a secure, decentralized data and reasoning infrastructure for autonomous aviation incorporating zero-trust architectures, edge computing, decentralized data exchange with transitive delegation for cross-organizational trust, and 4D geospatial databases to support real-time ML-driven coordination in the national airspace. This work underpinned NASA’s Unmanned Aircraft System Traffic Management (UTM) program and informed distributed architectures for autonomous space systems, including the Starling CubeSat constellation.

He also led development of AMPLIFY, an internal NASA platform for agentic AI with human-in-the-loop oversight, applied to mixed-autonomy air traffic management across the national airspace. AMPLIFY addressed how AI agents can assist human operators in coordinating autonomous and piloted vehicles sharing complex airspace, requiring the system to reason about incomplete state, communicate decisions transparently, and defer to human judgment under uncertainty.

This work was recognized with NASA’s Early Career Achievement Medal for pioneering secure, real-time data management systems, advancing NASA’s capabilities in mixed-autonomy traffic management and autonomous space missions.

Compound AI Systems for Machine Learning Engineering

Moustafa’s current research at Disarray investigates how compound architectures can enable autonomous ML engineering, applying the orchestration principles developed across his prior work to the problem of building ML models end to end. The prevailing approach to improving ML agents has focused on improving the underlying language models through reinforcement learning, test-time compute scaling, and growth in model and data scale. Moustafa’s work takes a different, complementary approach: rather than relying on model improvements alone, he investigates how semantic knowledge graphs, multi-agent orchestration, persistent memory, and domain-specific tooling can make existing models dramatically more effective at ML engineering tasks.

At Disarray, this research has produced a compound AI system that orchestrates the full ML lifecycle through several coordinated components: a multi-agent framework with inner execution and outer governance loops, a semantic knowledge graph constructed from heterogeneous data assets and source code via static program analysis, hybrid retrieval that respects graph topology and data provenance, persistent memory with quality-aware curation, and multi-model routing that matches subtasks to appropriate models. The knowledge graph is central, it provides agents with structured understanding of data semantics, code dependencies, and organizational context, addressing the context bottleneck that limited autonomous systems in each of Moustafa’s prior research domains.

This system achieved a 77.78% medal rate on MLE-bench, an evaluation of 75 real-world ML engineering competitions developed by OpenAI, representing the highest score recorded as of February 2026. The result suggests that compound architecture design can be as important as model capability for complex ML engineering tasks.

Collaborations

NASA Ames & University of Michigan (with Jiasi Chen, Abraham Ishihara, Kalmanje Krishnakumar): Data & Reasoning Fabric for autonomous aviation, mixed autonomy traffic management, zero-trust edge computing architecture

UC Berkeley RISELab (with David E. Culler, Ion Stoica, Joey Gonzalez, Joseph Hellerstein, Randy Katz, Raluca Ada Popa): Serverless computing across cloud-edge continuum, WAVE decentralized authorization (USENIX Security 2019), XBOS-DR smart building platform deployed across 100+ commercial buildings

Rutgers University & IBM Research (with Manish Parashar, Kirk E. Jordan): CometCloud distributed infrastructure, multi-cloud container orchestration, and HPC-as-a-Service

Large-scale science: Rutgers Cancer Institute (medical image analysis), UCSD Supercomputing Center (wildfire prediction), UT Austin (oil reservoir simulation), and Lawrence Berkeley National Lab (grid integration).