LLM serving · RAG · agent systems
Building & operating production AI stacks · ECARX · Suzhou, China
I ship and operate production LLM / RAG / agent systems — self-hosted Dify ops, retrieval pipelines, and upstream bugfixes in widely used open-source projects (Dify, RAGFlow, Qwen Code, Alibaba MNN, ModelScope ms-swift, QwenPaw, SAG, AG2, AgentScope, Promptfoo, Qingjian, FlashRAG, Qdrant, and related tooling).
Prefer fixes with reproducible RCA + tests, not drive-by docs.
| Project | What it does |
|---|---|
os-hybrid-kit  |
Portable Python library for OpenSearch hybrid search: BM25 + kNN fused with RRF or weighted score normalization, with thin Dify, LangChain and LlamaIndex adapters. |
dify-skills  |
Executable runbooks (SKILL.md Agent Skills) for operating self-hosted Dify Community 1.17: real HTTP routes, auth, payloads and failure modes. |
DocForge  |
LLM-powered document generator: write from a topic or match the style of an existing document, then export clean DOCX. |
If one of these saves you time, a ⭐ helps others find it. Issues and PRs are welcome.
Upstream contributions · LLM serving · RAG · agent runtimes
vLLM · Dify · RAGFlow · Qwen · MNN · ms-swift · Milvus · AgentScope · AG2 · Promptfoo · Qingjian · FlashRAG
| Area | Role |
|---|---|
| langgenius/dify | Contributor — API / Web / Agent runtime bugfixes (×8 merges) |
| ag2ai/ag2 | Contributor — knowledge store paths and logs, eval suite loading, skills frontmatter, A2UI fences, network graph loading, falsy Variable/Inject values, binary token sizing, shell sandbox ignore-list paths, A2A empty parts, cross-provider file IDs, empty knowledge dirs after delete (×14 merges) |
| QwenLM/qwen-code | Contributor — serve boot / MCP discovery / ACP transport reliability (×5 merges) |
| infiniflow/ragflow | Contributor — Agent retrieval / Wiki indexing (×2 merges) |
| Zleap-AI/SAG | Contributor — PDF/Office read fix / self-hosted MinerU 4.x integration / image, page and heading sidecars / hierarchical outline (×4 merges) |
| RUC-NLPIR/FlashRAG | Contributor — refiner model resolution / dataset attribute errors / GAOKAO-MM answer scoring (×3 merges) |
| alibaba/MNN | Contributor — converter fail-fast / TFLite index + createUnit (×2 merges) |
| modelscope/ms-swift | Contributor — distributed eval / generate path |
| agentscope-ai/QwenPaw | Contributor — tool-call coordinator reliability |
| agentscope-ai/agentscope | Contributor — app scheduler start-time anchoring / local file URL decode / memory frontmatter fence at EOF (×3 merges) |
| promptfoo/promptfoo | Contributor — eval var substitution / CSV similarity assertion types (×2 merges) |
| qingjian-team/qingjian | Contributor — macOS candidate-window font fallback for unreadable outline formats |
| qdrant/qdrant-client | Contributor — local-mode validation parity for score_threshold, prefetch and fusion queries / formula geo points (×2 merges) |
48 merged upstream PRs so far.
More work in flight across Qwen / agent / RAG ecosystems (agentscope, LightRAG, Dify, docling, qdrant).
Thanks for stopping by — happy to chat about LLM infra, RAG, and agent systems.