Corey JamesI'm a systems architect, engineer, and independent AI researcher whose work spans GPU computing, high-performance inference, autonomous agents, and enterprise AI architecture. My career has taken me from designing complex enterprise platforms and Microsoft ecosystems to developing inference engines, GPU memory architectures, model optimization techniques, and the infrastructure required to support large-scale autonomous AI systems.
I work across Rust, C++, CUDA, Python, and C#, developing software for NVIDIA and Intel accelerators while maintaining the broader architectural perspective that comes from years of enterprise technology experience. I am equally comfortable discussing identity, governance, and enterprise integration as I am investigating GPU memory bandwidth, quantization, inference scheduling, and kernel performance. What interests me most is understanding why systems behave the way they do, particularly when accepted architectural assumptions impose limitations that may not actually be necessary.
From enterprise architecture to AI systems research
My foundation is in enterprise systems architecture, where I have spent years designing and delivering technology that must operate reliably within complex business environments: Microsoft 365, Azure, Copilot, identity and security, automation, and custom AI development. It taught me to approach technology as a complete operational system rather than a collection of interesting components.
As AI matured, my work moved progressively deeper into the technology itself. Integrating models into business processes led to building agent orchestration infrastructure, which led to investigating how models manage context, access memory, execute inference, and share computational resources. Today, a substantial part of my engineering and research focuses on the layers beneath the applications, where architectural choices directly affect performance, capability, and cost. I believe the next generation of AI systems will depend as much on advances in memory management, inference infrastructure, and execution architecture as on improvements to the models themselves.
Research, engineering, and original systems
I develop my own systems to investigate problems that existing platforms do not adequately address. Two examples are AgentOS, an autonomous engineering and work execution platform, and ActiveKV, an alternative memory architecture for large language model inference.
AgentOS grew out of a fundamental question: how can independent agents perform substantial engineering work without sacrificing control, correctness, accountability, or reproducibility? Its architecture incorporates work-order compilation, governed execution, persistent memory, parallel agent coordination, and evidence-based verification, so that autonomous work is a controlled engineering process rather than an open-ended conversation with a model.
ActiveKV approaches a different problem at the infrastructure level. Conventional inference serving depends heavily on GPU-resident key-value caches, creating memory constraints that directly affect context capacity and concurrency. My work investigates host-backed KV memory, compression, data movement, native execution paths, and scheduling to change that relationship while preserving correctness and competitive performance.
I conduct this work on my own multi-GPU research infrastructure, including NVIDIA RTX and Intel Arc Pro accelerators, which lets me investigate architectural differences, test implementation choices, and measure system behavior independently. These projects are part of a broader effort to understand how AI systems can be redesigned when the assumptions embedded in existing infrastructure are treated as engineering questions rather than permanent constraints.
Engineering with autonomous AI
My approach begins with understanding the underlying problem before choosing an implementation: why something fails, where the actual bottleneck is, and whether improving the current design is more appropriate than replacing it. I separate correctness from performance. An optimization that produces an impressive benchmark but compromises intended behavior, consumes excessive resources, or cannot be reproduced is not an advancement. I establish baselines, instrument the systems I develop, evaluate competing approaches, and use measured results to decide which changes survive.
That discipline is what makes my engineering methodology around autonomous coding agents work. Rather than using AI as an interactive coding assistant, I architect and direct coordinated groups of agents that perform implementation, investigation, testing, and review under explicit technical constraints and acceptance criteria. The implementation of much of my recent software has been performed by AI coding agents across multiple frontier and open-source model families. I design the systems, set the architectural requirements, direct the technical decisions, define the validation, and govern how work is accepted; independent reviewers, automated verification, and evidence requirements distinguish completed engineering from plausible-looking output.
This has let me develop and maintain systems of hundreds of thousands of lines while concentrating my own effort on architecture, research, difficult technical decisions, and performance engineering. It is also a practical environment for studying autonomous AI itself: the limits of model reasoning, long-running execution, context management, delegation, and verification. I see it as an emerging engineering discipline rather than a productivity technique.
Where I'm focused today
My current interests sit at the intersection of high-performance computing, model systems engineering, and autonomous software development: GPU-native inference, efficient memory architectures, heterogeneous acceleration, model optimization, and the infrastructure necessary to operate autonomous agents at meaningful scale. I continue to bring enterprise architecture experience into this work because advanced AI capabilities ultimately need to function within real operational environments. Performance without reliability has limited value, just as powerful autonomous systems without governance are difficult to trust.
Contact
Sent. I read every message.