Company
Projects
GGUF Token Decoder
Turns a local GGUF model's raw token IDs back into text using only the model file.
Transactional Relational Database Engine
TRDE, a database engine I wrote in pure Python with no dependencies: a pager, B+tree indexes, crash-safe transactions, a SQL parser and a query planner.
HisokaOS
A 32-bit x86 operating system I wrote from scratch in C and assembly. It boots its own kernel in QEMU to an interactive shell, with paging, PCI, an RTL8139 network driver and a RAM filesystem.
All projects
- The Lazy DriverSelf-driving perception and path prediction on dashcam footage.
- BillA local conversational system with memory, retrieval and fact checking.
- Reachy RLReachy 2 in MuJoCo, set up for reinforcement learning.
- Editor BrowserA full browser tab inside the VS Code editor.
- App InstallerA Windows application with an NSIS installer.
Other projects
Flagger
The traffic proxy in front of my GPU server's local language models. It holds Ollama's port and runs one GPU job at a time, most important first: Webbify clients, then public chat, then my own tools, then background work. It also checks every Webbify service every 15 seconds and restarts any that fail three checks in a row.
Ducky Studio
My desktop app for working with the models on my GPU server. It has chat with tool calling, scheduled tasks that run against a real model, model downloads with live progress, and a view of both machines.
Latent Diffusion Image Generator
An image generator on a 6 GB consumer GPU. I instantiate the full latent diffusion architecture myself (U-Net, VAE, CLIP text encoder, DDIM scheduler) and load open pretrained weights into it. It loads on demand and unloads itself when idle.
Motion-Adapter Video Generator
Short, low-resolution video clips on the same 6 GB GPU, using a motion module added to a Stable Diffusion U-Net. I tracked an older checkpoint that only ever output noise to a fault in how its time layers merge, and switched to AnimateLCM.
On-Demand Language Model with Web Search
A 4-billion-parameter model (Qwen3-4B) served on demand. It looks things up with a live web-search tool when an answer is time-sensitive.
Wingwall Neural World
Real driving scenes rebuilt as navigable 3D worlds with Gaussian splatting, made as a place to test a self-driving planner. It reaches 21.5 dB PSNR on held-out views.
Gallery
Lidar Map in CARLA
HisokaOS setup screen
Ducky Diffusion
