Thanks for the kind words.
To answer your question:
I architect everything. Every system design, every architecture doc, every interface contract, every integration decision. The kernel is a from-scratch Rust kernel. The hypervisor is a clean-room Rust implementation. The security model, the cognitive architecture, the 10-layer enforcement stack. All of that comes from 20 years of building systems and an autistic brain that can hold all of it in memory simultaneously.
I then orchestrate upto 10-12 parallel AI coding sessions, each working on a separate component with strict file ownership and interface contracts defined upfront.
It’s closer to being a lead architect directing a development team than it is to prompting ChatGPT and hoping for the best.
You can’t vibe code a kernel. You can’t vibe code a hypervisor. You can’t vibe code a BPF LSM policy compiler.
These things require deep understanding of what they are and how they connect. The AI types fast under my direction.
I’ve developed a methodology around this: roughly 25% of the codebase is architectural markdown that guides code generation, proto files coordinate between sessions, integration happens between batches.
It’s a disciplined process, not random prompting. Nobody accidentally produces 1.3M lines that compile and pass 3,900+ tests (that's the number of tests for the core project only, not the entire 1.3M LoC)
To answer your question:
I architect everything. Every system design, every architecture doc, every interface contract, every integration decision. The kernel is a from-scratch Rust kernel. The hypervisor is a clean-room Rust implementation. The security model, the cognitive architecture, the 10-layer enforcement stack. All of that comes from 20 years of building systems and an autistic brain that can hold all of it in memory simultaneously.
I then orchestrate upto 10-12 parallel AI coding sessions, each working on a separate component with strict file ownership and interface contracts defined upfront.
It’s closer to being a lead architect directing a development team than it is to prompting ChatGPT and hoping for the best.
You can’t vibe code a kernel. You can’t vibe code a hypervisor. You can’t vibe code a BPF LSM policy compiler.
These things require deep understanding of what they are and how they connect. The AI types fast under my direction.
I’ve developed a methodology around this: roughly 25% of the codebase is architectural markdown that guides code generation, proto files coordinate between sessions, integration happens between batches.
It’s a disciplined process, not random prompting. Nobody accidentally produces 1.3M lines that compile and pass 3,900+ tests (that's the number of tests for the core project only, not the entire 1.3M LoC)