what
Below is a loose map of current projects, older experiments, and recurring obsessions.
These are the monsters I keep walking into: CUDA math, agent infrastructure, factory floors, Gaussian splats, and other high-entropy explorations.
-
Gaussian Moat Cuda
Computational number theory project around the Gaussian Moat Problem: GPU-first search, verification, and paper-grade evidence.
Current result: a novel CUDA-backed approach with custom kernels and roughly 20k per dollar speedup on a single 4090 compared to the previous baseline. The repo is public; the paper is coming soon.
-
Factory AI Consulting
Practical factory AI work: knowledge-base agents over messy documents, OCR, custom harnesses, benchmarking, messenger integrations, and deep local-model research for on-premise deployment.
The interesting part is not "chat with PDF". It is making scattered scans, regulations, amendments, and department lore answerable with sources, while pushing local models toward frontier-quality behavior on Mac Mini / NVIDIA-class hardware.
-
Tortuise
Terminal-native 3D Gaussian Splatting viewer: real `.ply` and `.splat` scenes rendered as symbols in a terminal!
CPU-first Rust, six render modes, no GPU required, and a surprisingly alive proof that terminal graphics can still be a frontier, even over SSH to potato! (Jetson Orin Nano)
-
Pratchett-OS
Private knowledge, memory, and coordination layer behind the public work: notes, agents, sessions, tickets, skills, and writing flow.
Not intended to become public as a repo, at least for now. Some mechanics may become posts once they are clean enough to explain without leaking the house wiring.
-
Tennis XGBoost Autoresearch
Karpathy-style autoresearch loop on tennis, now continuing as a broader XGBoost playground with Minecraft speedrunning / MCSR Ranked.
-
agent-tickets
Filesystem-native ticket engine for agent work: markdown cards, initiative folders, explicit lifecycle, dependencies, and optional agent-mux dispatch.
Cards are the database, git history is the audit trail, and the state machine is the guardrail. Small enough to paste into an agent, structured enough to automate.
-
tg-useragent-cli
Agent-first Telegram user-account CLI in Go: local SQLite mirror, scoped exports, dry-run confirmations, and a safety model for letting agents work with real Telegram context.
-
tg-agents-wrapper
Telegram bot wrapper for Claude Code and Codex: mostly a toy, but also the mobile surface where Macupos learned to be useful.
-
Hardware Lab
Physical-world experimentation thread behind the factory and splat work: Bambu A1 printing knowledge, Jetson Orin Nano edge AI, and apartment-to-splat capture.
Mostly not one public repo. More like a lab bench with receipts: 15kg of printing experience harvested into settings, NanoOWL running at 30+ FPS on Jetson, and a local phone-video-to-splat pipeline.
-
wet claude
Claude Code context optimization proxy. Useful proof-of-concept for surgical tool-result compression; not currently active because native hooks are the cleaner path.
-
Fieldwork Skills
Public collection of operational skills for AI coding agents. Mostly legacy now, but it captured the moment where skills became judgment in folders.