Systems engineer — Rust, GPU/CUDA, ML infrastructure. Low-level and performance-first, with 15+ years across hardware, firmware, and software. Based in Kostanay, Kazakhstan; open to remote work or relocation.
I spent years close to the metal — repairing electronics, writing bare-metal firmware, designing PCBs — and now build systems software in Rust: a from-scratch ML inference stack and a GPU-rendered GUI framework, both without a Python runtime.
synthos — a local AI desktop studio (~143k lines): an agentic chat that runs 27B–125B models from single-file bundles, a block-based notes workspace (canvas, kanban, Gantt, mind maps, calendar), a node editor for video / music / speech generation, a code editor and a local knowledge base. Everything on-device, no Python. MIT / Apache-2.0.
synaptix — a native Rust engine for running and training models (~232k lines, 1800+ tests): an alternative to the Python ML stack, no PyTorch, no Python runtime. Hand-written CUDA kernels via NVRTC with NVFP4 / MXFP8 quantization, checked bit-for-bit against reference implementations. Gemma-4 26B at 210 tok/s decode, a 125B MoE with a 262k context on a 24 GB laptop card. MIT / Apache-2.0.
syngui — a retained-mode GUI framework in Rust (~166k lines). wgpu rendering, an own text stack (shaping, glyph atlas, color emoji), a CSS-like styling language with cascade and animations, 110+ widgets (terminal, code editor, block document editor, charts, tables, video). Runs on desktop, Android, and WebAssembly. MIT / Apache-2.0.
Earlier commercial work: a central-heating controller with a custom architecture and GSM-OTA bootloader (in production for years), and a Qt/C++ SPI-flash recovery tool used daily in an electronics-repair shop.
Rust · C++ · Assembly (x86 / AVR / PIC) · CUDA (NVRTC runtime kernels, GEMM/GEMV, fused kernels, NVFP4 / FP8 quantization, Nsight profiling) · wgpu / WGSL · embedded (STM32, AVR, PIC; bare-metal, DMA, bootloaders) · Linux (Arch) · FFI · KiCad
Self-taught across the stack. I verify claims instead of trusting them — I benchmark and profile my own work and report the real numbers, including where it loses.
Email — vitamindbnfkz@gmail.com
If any of this is useful to you, you can support the work via PayPal.