The lab.
Deep dives, project stories, and engineering notes - anything we build that's worth writing about.
Append-Only Storage and Two-Tier VACUUM: How KAI Handles Deletes Safely
Borrowing from PostgreSQL and LSM-tree databases, KAI's document store achieves zero-copy reads, crash-safe deletes, and background physical reclamation.
Training-Free Vector Quantization: Lloyd-Max Codebooks on the Unit Sphere
How KAI compresses 6 KB float32 vectors into ~768 bytes - without ever looking at your data - using the geometry of high-dimensional spheres and optimal scalar quantization theory.
Inside the SIMD Search Pipeline: How KAI Scans Millions of Vectors in Milliseconds
A deep dive into nibble-split lookup tables, blocked memory layouts, and architecture-specific SIMD kernels - the techniques that make sub-millisecond vector search possible.
Parsing BMP Files on a Microcontroller With Almost No RAM
How to read, decode, and draw Windows BMP images on a tiny microcontroller by streaming pixels straight from an SD card - no framebuffer, no megabytes, just a couple of rows of RAM.
Bridging Async and SIMD: The Tokio + Rayon Architecture of a High-Performance Rust Server
Building a vector search server that handles hundreds of concurrent HTTP requests while running SIMD operations at full CPU efficiency.
WebAssembly Is the Plugin System We Always Wanted
WASM finally cracked the oldest problem in software extensibility: running someone else's untrusted code, written in any language, safely and fast. Here's why plugin systems everywhere are quietly being rebuilt on it.
Settling the Framework Debate: Why Svelte 5 Won
After fifteen years of religious wars over frontend frameworks, one of them quietly solved the problem. A technical case for Svelte 5 - runes, the compiler, the benchmarks, and the honest tradeoffs - from a consultancy that ships across the stack.