◆ Systems · inference · agent tooling
CPU inference engines in C99, SIMD kernels, and the agent tooling that drives them. Systems developer from Nepal; I lead PulseOnix.
01 — The collection
Dot product, L2 and cosine, hand-written across SSE2, AVX2, AVX-512 and NEON. One binary picks the widest path the CPU has. Zero dependencies, stable C ABI.
Everything is a plugin — kernels, architectures, tokenizers, samplers, even the generation loop. Swap any of it at runtime by config. GGUF, OpenAI-compatible API, 21 test suites.
Hidden rules learned from experience alone — no corpus, no pretraining. From tabular world models to self-modifying exploration across six phases.
One CLI and one MCP server for every HF account, repo and Space — built for humans and agents alike. Installable via pipx.
02 — Run it
vek's dot product, compiled to WebAssembly twice — plain f32 against SIMD128 — and raced on the machine you're holding. Nothing is uploaded.
Four sizes, plain f32 against SIMD128. About a second.
Median of three runs per size, warmed up first; 2 flops per element. This is WebAssembly in your browser, not a native build: a browser exposes one vector path, so two of vek's four tiers are reachable from here. The native numbers are in the repository.
03 — The bench
Six ecosystems, one habit: understand the layer below before trusting the one above.
C · C++ · Rust · Zig · Go · Assembly · SIMD · cache-aware design
Python · PyTorch · inference · quantization · RAG · RL · agents
TypeScript · React · Astro · Three.js · WebGL · Node · CSS
ESP32 · Raspberry Pi · Arduino · SystemVerilog · I²C / SPI
Linux · Git · Docker · CI · Cloudflare · MCP · Ghidra · GDB
design systems · typography · motion · accessibility · Figma
