Hi, I'm Guishi Wang. I build practical AI systems across LLM applications, retrieval, agent workflows, document intelligence, and computer vision. I care less about isolated demos and more about turning model capabilities into systems that are reusable, maintainable, and verifiable. I also contribute focused fixes to open-source projects, taking issues from root-cause analysis through regression coverage and compatibility checks.

• 🤖 Building LLM applications, model gateways, and agent workflows
• 🔎 Exploring Chinese semantic retrieval, reranking, and document intelligence
• 👁️ Applying computer vision to object detection, similarity, and image understanding
• 🧱 Engineering for reliability, observability, and clear system boundaries
• 🛠️ Contributing regression-tested fixes across AI tooling, runtimes, compilers, databases, and developer infrastructure

AI Applications
   │
   ├── LLM Gateway       routing · rate limits · resilience · audit
   ├── Knowledge & RAG   embeddings · retrieval · reranking
   ├── Agents & Skills   tool use · workflows · reusable capabilities
   └── Vision & Docs     parsing · similarity · object detection
   │
   ▼
Engineering Systems
   │
   ├── Services          FastAPI · Next.js · Java
   ├── Data & Runtime    MySQL · Redis · vector stores · Docker
   └── Operations        observability · security · maintainability
   │
   ▼
Applied Scenarios
       knowledge work · procurement & compliance · developer productivity

• Guishi-Skills — A growing collection of reusable Codex skills for data analysis, publishing, presentations, research, and codebase understanding.
• model-switch — An OpenAI-compatible LLM gateway with multi-provider routing, rate limiting, circuit breaking, queues, auditing, and an admin console.
• terms-rag — An offline RAG system for Chinese tender clauses, powered by BGE embeddings, reranking, ChromaDB, and FastAPI.

Lately, I've been contributing focused fixes across AI tooling, runtimes, language tooling, databases, and developer infrastructure. I usually start from the root cause, keep the implementation narrow, and validate the concurrency, compatibility, security, or performance edge cases around it.

• Distributed reliability & media pipelines — implemented Redis-leased global task slots, atomic repair of corrupted video caches, and consistent enum defaults in MoneyPrinterTurbo (#1742, #1743, #1744); fixed Electron edge cases around failed upload streams and pre-aborted message boxes (#55520, #55519).
• AI model & runtime compatibility — improved InvokeAI calibration retry isolation and LoRA/LyCORIS handling for DoRA, Diffusers Anima, and factorized layouts (#10077, #10075, #10076, #10060); fixed Codex routing, reasoning-effort, output-token, and usage-metering behavior in cc-switch (#8998, #8999, #9001).
• Compilers, SDKs & type systems — corrected async policy chains and emitter option/server parameter handling in TypeSpec (#12448, #12447, #12446); fixed cross-namespace models, binary response types, and nullable enum generation in Kiota (#8517, #8516, #8515); and repaired narrowing and generator-expression behavior in Pyright (#11952, #11951, #11950).
• Developer tools & data platforms — isolated ECC observer scopes, removed duplicate Git status work, and reduced hook false positives (#3817, #3818, #3819); added Unicode SQL, wrapped ClickHouse type, and PostgreSQL partition fixes to DBeaver (#42637, #42638, #42639); and contributed targeted fixes to Playwright, RustDesk, Neovim, and VictoriaMetrics.

View all pull requests →

If you are interested in LLM engineering, RAG, agents, document intelligence, or computer vision, feel free to connect and exchange ideas through GitHub.