🌟 I ship one runnable open-source AI project every day. AI Daily Hub indexes 70+ zero-dependency frameworks (sparse MoE · decision models · embeddings · world models · agent orchestration · video diffusion), each with the underlying principle explained and reusable code. 每天一个 AI 热点，一个「填配置就能跑」的开源项目。
I build control planes for AI infrastructure — the layer between your application and the models it depends on. My work starts from one observation: most AI projects fail in production not because a model is weak, but because governance, observability and reliability were bolted on afterwards. So I build systems where those three are kernel primitives, not plugins.
Everything here is open source, self-hostable and model-agnostic.
Three of these projects are the same idea applied at three layers:
| Layer | Project | The primitive it makes native |
|---|---|---|
| Text | unified-ai-gateway |
One OpenAI-compatible endpoint in front of many LLM providers — routing, quotas and cost become infrastructure concerns. |
| Video | video-forge-studio |
One REST API in front of six video engines — queueing, retries and multi-tenancy become infrastructure concerns. |
| Agents | loopforge |
One governed loop kernel for long-running agents — checkpointing, budget limits and audit become infrastructure concerns. |
The other three close the loop: you cannot govern what you cannot measure, so
argus-eval handles evaluation and tracing, moe-bench-studio handles model
selection, and ai-daily-agent keeps the whole portfolio moving.
unified-ai-gateway —
Control plane for the text layer.
A self-hosted, OpenAI-compatible gateway with smart routing, load balancing,
failover, rate limiting and cost tracking. Swap providers without touching
application code. TypeScript · pnpm monorepo
video-forge-studio —
Control plane for the video layer.
Six generation engines (self-hosted + commercial API) behind one REST API, with
Celery queues, retries and quotas. A built-in Mock engine means the whole
pipeline and its CI run with zero GPU and zero API keys.
Python · FastAPI · Celery · 51/51 tests
loopforge —
Governance kernel for long-running coding agents.
Runs autonomous software-engineering loops over hours to days with checkpoint /
resume, budget guardrails and a full audit trail. Recoverable, governable,
auditable, observable. Python · FastAPI · React
argus-eval —
Evaluation and observability for AI agents.
Tracing SDK, span and score ingestion, 15+ scorers, dataset management and
experiment comparison, on a self-hosted dashboard. Python · OpenTelemetry
moe-bench-studio —
Inference and benchmarking for trillion-parameter MoE models.
Ten benchmark suites, multi-engine inference and a quantization planner, so
model selection is a measurement rather than a guess. Python · FastAPI · vLLM
ai-daily-agent —
The automation behind the portfolio.
Discovers daily AI topics, generates a project, publishes it to GitHub and
monitors the metrics — including a REST-API transport layer that keeps working
in networks where git push is blocked. Python
- Self-hostable by default. No mandatory SaaS in the critical path.
- Model-agnostic. Providers are configuration, never an architectural commitment.
- Runnable on a laptop. If a newcomer needs a GPU to see it work, the packaging is broken — hence offline mock engines and inline queue fallbacks.
- Governance as a primitive. Quotas, audit trails and budget limits belong in the kernel, not in a wrapper.
- 💬 有问题 / 想聊：开 Discussion 或 Issue
- ⭐ 喜欢就 Star，让我知道有人在用