secure-shell / session 01
kernovex / Tim
Kernovex / Tim / software developer

I build software I want to use.

I’m Tim. I go by kernovex. I’ve built my own control panel, worked on a CPU-only inference engine called OwnCPU, and written little tools for the jobs I get tired of doing by hand.

kernovex:~$./initialize
interactive index

Ask the terminal.

Try about, projects, systems, story, or contact. Keyboard-first by design.

portfolio-cli — connected
Use Tab to focus the prompt. Type help for the command list.

Software I’ve built

OwnCPU — CPU-only inference

I wanted to run a serious model on the machine I actually own. That turned into C++ work on quantization, NUMA placement, and compute kernels, with correctness checks after each change.

02 / WHEN IT IS SLOW

Kernel experiments

One change at a time: memory bandwidth, matmul shape, cache locality, fused paths. I keep the bad results too. They usually explain the good ones later.

profilingBLASbenchmarks
03 / THE MACHINE

Linux as a workbench

Long-running jobs, systemd units, process isolation, logs, and the small bits of care that keep an experiment from disappearing overnight.

Linuxsystemdinfra
04 / SMALL TOOLS

Tools I wish I had

A pile of little scripts, CLIs, and interfaces that make the current problem easier to see. Most start as a shortcut and become useful after the third rewrite.

PythonNode.jsCLI

This portfolio

I wanted visitors to be able to type into the terminal as well as browse the page. The commands work in the browser, and the site runs as separate HTML, CSS, and JavaScript files on my EPYC server.

A bit about me

A lot of what I like about writing software is the same stuff I like about building PCs. I want to know why a machine won’t boot, what a BIOS setting actually changes, and what the manual left out.

With OwnCPU, that meant following inference down through C++, tensor layouts, quantization, memory pressure, Linux scheduling, and the gap between “it runs” and “it runs well.” I learned more from the wrong turns than from the clean benchmarks.

I try to write down what changed, what failed, and which measurement I still don’t trust. It saves me from rediscovering the same bug a week later.

current modedeep systems work
favorite constraintmake it explainable
hardware biasCPU-first, practical
default loopmeasure → change → verify
next horizonmore useful tools
open channel

If you’re building close to the metal, we’ll probably have something to talk about.

I’m always interested in thoughtful systems work, useful tools, and performance problems that refuse to be hand-waved away.

open the terminal