• Large-scale Inference Scheduling: Optimizing GPU utilization across clusters.
• Multi-model Routing: Intelligent traffic governance for heterogeneous models.
• Production LLMOps: Moving from "demo" to "day-2" operations.
• Cloud-Native Abstractions: Defining standard APIs for AI workloads.

• semantic-router (Committer, Current Focus): Defining the decision-making layer for multi-model LLM serving.
• llm-d: Building the cloud-native infrastructure for disaggregated LLM inference.
• HAMi & Kueue: Used in Kubernetes-native batch scheduling & GPU virtualization.
• Istio: Standardizing traffic governance for service mesh.
• Karmada & Kubernetes: Essential for multi-cluster orchestration at scale.

I ship solo via a multi-worktree, LLM-assisted workflow — gather → split → isolate → ship → loop. The full set of Claude Code skills I use daily lives in ./skills.

I build tools to fix my own problems.

• Chrome TabBoost: Browser tab overload is a bug. I patched it with an extension.
• MacMusicPlayer: A minimalist, clean music player for macOS.
• ConfigForge: Manage ~/.ssh/config and kubeconfig without the headache.
• gmc: AI-powered Git commit messages.
• mdctl: AI-powered Markdown workflow automation.
• hf-model-downloader: Painless Hugging Face model downloads.
• LogoWallpaper: Generating brand assets shouldn't take 30 minutes.
• SaveEye: A minimalist eye care reminder that doesn't annoy you.
• homebrew-tap: I deliver binaries. I don't just dump code. (brew tap samzong/homebrew-tap)
• mirrormate: Docker pulls failing? I fixed it with magic.
• swagger-online: Unified Swagger UI. No more tab chaos.
• ai-icon-generator: I needed icons, so I built a generator.
• Recall: I lose great AI answers weekly. A local-first TUI to search my Claude Code, Codex, and OpenCode history.
• gofs: python -m http.server is slow. Rewrote it in Go.
• GithubNotifier: Checking PRs in browser tabs is a ritual. Moved it to the menubar.
• gh-x: Batch repo operations as a gh CLI extension. One-by-one doesn't scale.
• convostore: Stateful LLM APIs shouldn't be this hard. A Redis-backed conversation store for OpenAI Responses API / vLLM.
• prompts: Prompt engineering is the real frontend of LLMs. I track mine here.
• mote: Rewrite any selected text on macOS via Markdown-defined commands.
• merge-scout: A Claude Code skill for vibe coding — ranks GitHub issues by contributability × merge probability so your agent picks work that will actually land.
• openclaw-gateway-tunnel (RFC, design phase): Zero-binary ngrok tunnel for OpenClaw gateways. Looking for co-designers before code lands.
• moltbot-channel-feishu (legacy, superseded by OpenClaw official): Feishu/Lark channel plugin for Moltbot/Clawdbot.