chen hao

You can fine-tune LLMs (LoRA, DPO, GRPO, RLHF), run distributed pretraining with FSDP2 and Megatron-LM, serve models via vLLM or TensorRT-LLM, build RAG systems on Chroma/Qdrant/Pinecone/FAISS, perform OSINT and web pentesting, build DCF/LBO Excel models, and generate images, video, and music.

ML engineers running training and inference pipelines, security researchers doing penetration testing and supply-chain forensics, investment analysts building valuation models, and creative producers generating video, comics, and music. Most skills assume comfort with Linux/macOS terminals and Python environments.

Skills are loaded as SKILL.md folders with YAML frontmatter inside a GPUCLOUD-compatible agent runtime. Dependencies are declared per skill (e.g., torch, transformers, vllm) and installed via pip; CLI-based skills require commands like curl, jq, or gh verified in the runtime first.

Nearly all skills carry MIT licenses, with a few Apache-2.0 (finance modeling, hyperframes) and one proprietary PowerPoint skill. Many rely on free tiers or keyless services (DuckDuckGo, SearXNG, Solana RPC), while others need API keys such as USDA, Notion, Shopify, or 1Password service accounts.

Core requirements include Python 3, curl, and platform support for Linux/macOS/Windows. ML skills need PyTorch 2.0+, transformers, and often CUDA GPUs; specific skills require Blender 4.3+, ComfyUI, Docker, Node.js with ffmpeg, or credential files like Google OAuth2 tokens and Shopify access tokens.