essays and notes on vector databases, simd, allocators, and tiny vms.
This step was years of proving an idea, not years of polishing a finished product. The bet was that vectors, reactions, and grammars could share one dispatch.
The initial design thesis, the K influence, and the first rules for data layout, parsing, and concurrency.
8 posts
The value of ThePlatform was not any one subsystem. It was that one runtime, one debugger, and one extension model survived production pressure.
Proofs, retrospectives, and the production lessons that made the unification claim credible.
7 posts
This was not a side project for optics. It was a practical detour where fairness, wakeups, and panic safety had to be learned the hard way and in public.
The MPMC channel work that sharpened the concurrency discipline behind later systems work.
4 posts
Rayforce began as a deliberate restart in C because embeddability mattered more than implementation comfort. The constraint, not the language preference, drove the move.
The transition away from a server-shaped Rust system toward an embeddable C library and its language core.
6 posts
The hard part was not inventing kernels. It was making the system measurable, bindable, allocatable, documented, and honest enough for other people to use.
Allocator work, binding pressure, benchmarks, release work, and user-facing product decisions.
6 posts
AxlDB was an honest negative result. The experiment mattered because it showed Rayforce was not missing a secret cleaner core.
The side experiment that clarified where the engine ended and the database began, then got shelved.
2 posts
This step is about giving up the comfort of separate projects. The language users knew and the engine I wanted had to become one system.
The merge, the morsel pipeline, spec transcription, and the new query surfaces unlocked by the new engine.
4 posts