📍 Ship fast. Interate faster.

Deep in coding mode — training models, orchestrating agent swarms from my phone, and building games where NPCs think for themselves. Ship fast, iterate faster.

• 📐 LLM-Oriented Design (LOD) Patterns — Design manifesto for the post-human-readability age. Three principles — Context Management, Feedback Loop, Tooling — that cut token consumption up to 40x. Drop the rules into your AI coding assistant and it writes LLM-friendly code by default.

• 🚀 Catapult — Ship checkpoints to production in one command, not just track them. Register, build Docker images (25+ templates), deploy with GPU awareness, benchmark latency, evaluate quality — one platform, one command. Comes with a Python SDK, MCP server.
• 🧪 Orze — Auto-research on autopilot. Runs the full loop: generate ideas → train on GPUs → evaluate → learn → repeat. Multi-agent, multi-machine, filesystem-coordinated. Python · PyTorch · Claude
• 🐂 oxRL — Lightweight post-training framework for LLMs and VLMs. GRPO, PPO, GSPO (MoE-friendly), DPO, and more — with Ray distributed rollouts and DeepSpeed ZeRO-3 training. Claude/Gemini/Codex friendly. Python · PyTorch · Ray · DeepSpeed · vLLM

• 🌍 polyworld.us private — Browser-based voxel sandbox game with LLM-driven NPCs (Gemini). Full multiplayer with real-time physics. React · Three.js · Colyseus · Redis · Gemini · Docker
• 🎒 polyworld-inventory private — Inventory management system for Polyworld. React · TypeScript · Vite

• 📡 Teleterm — Control remote Linux/macOS terminals via Telegram. Command your agent swarms from your pocket.
• 🐕 readog private — Chrome browser extension for reading and data extraction. JavaScript · Shell

• Training LLMs — Building models
• Vibe-coding games — NPCs that think, worlds that evolve, physics that crunch
• Multi-agent orchestration — Running Claude swarms with prompt frameworks and coordination protocols
• Shipping full-stack AI apps — From model training to production deployment in days, not months

ML/AI        → PyTorch, Qwen, InternVL, LLaVA, Gemini, Claude, CUDA
Backend      → Python, FastAPI, Go, C, Node.js
Frontend     → React, TypeScript, Three.js, Vite
Infra        → K8s,Docker, Redis, PostgreSQL, Nginx, Ceph
Automation   → Telegram bots, Claude Code, agent swarms

"Ship beats perfect" — I train models, build games, and manage it all from my phone. When the NPCs start writing better code than me, I'll know I've won.

• 🧙‍♂️ GitHub handle is warlockee — the warlock who ships code by moonlight
• 📱 Controls terminal swarms from Telegram — true pocket vibe coding
• 🤖 Runs multiple Claude instances concurrently like summoning familiars
• 🎮 Building a game where the NPCs are smarter than the players
• 🀄 Speaks English and Chinese

_{🧙‍♂️ This warlock is happy vibing}