Deep in the CUDA programming model — writing kernels, understanding warp-level execution, and slowly building intuition for how data movement dominates everything on the GPU.

Building a multi-agent CMS that coordinates LLM-backed specialists through a shared context layer. The goal: agents plan, edit, and publish collaboratively, human in the loop only when it matters.

Working through Programming Massively Parallel Processors (4th ed.) cover to cover. The exercises are brutal. Highly recommended.