AI engineer building multimodal systems, agent infrastructure, and reliable open-source software.
Computer Vision · AI Agents & RAG · Python Engineering · Full-stack Delivery
UESTC · China
I turn model capabilities into systems that can be tested, inspected, and shipped. My work connects vision models, LLM agents, retrieval, APIs, and user interfaces—with particular attention to failure handling, traceability, and maintainable engineering.
- Building multimodal inspection and agent workflows in Python.
- Contributing production fixes, tests, and typing improvements upstream.
- Comfortable owning the full loop: issue analysis, implementation, testing, review iteration, and delivery.
- Sharing AI engineering notes: 200+ followers and 1.5k+ likes on Xiaohongshu.
15 pull requests merged across 5 external projects. Selected work includes async cancellation, safer agent plugin installation, datasource edge cases, type safety, CI cleanup, MCP tooling, and document processing.
| Project | Contribution | Merged | Project |
|---|---|---|---|
| cloud-init | Datasource reliability, type-safety, and CI cleanup across CA certificates, Netlink, Bigstep, Vultr, RbxCloud, DigitalOcean, and the Ruff baseline | 8 PRs |  |
| fast-agent | Prompt child cancellation, contained skill/plugin installs, and collision-free stream captures | 3 PRs |  |
| wemake-python-styleguide | MCP server for explaining Python style violations | 1 PR |  |
| Railtracks | Paragraph-level PDF breakdown strategy | 1 PR |  |
| OpenVideoStudio | Cross-platform path handling and clearer Ollama connection diagnostics | 2 PRs |  |
- Propagate child cancellation promptly in an async agent runtime.
- Contain skill and plugin installs to prevent writes outside their target directories.
- Restore cloud-init's main-branch format checks with a focused Ruff cleanup.
- Handle metadata and IP edge cases in the Vultr datasource.
- Add an MCP violation-explanation server for Python lint diagnostics.
I value small, reviewable changes backed by focused regression tests and clear maintainer communication.
|
Evidence-first multimodal fire-safety inspection. VLM and YOLO outputs flow through specialist review, traceable history, and human feedback.
|
A JD-aware job-search workspace for resume diagnosis and generation, opportunity scoring, application tracking, and interview preparation.
|
|
A visual inspection system combining YOLOv8 detection with language-model reasoning for behavior and compliance analysis.
|
Retrieval experiments in which agents decide when to search, verify evidence, and refine their answers.
|
| See | Reason | Build | Ship |
|---|---|---|---|
| VLM, detection, OCR | Agents, RAG, tool use | Python APIs, Vue interfaces | Docker, Nginx, tests, tracing |
- Reliable agent runtimes: cancellation, isolation, tool/plugin boundaries, and observable execution.
- Multimodal inspection systems that keep conclusions connected to inspectable evidence.
- Evaluation and regression tests built from real failure cases rather than demo-only success paths.
Build systems that do more than predict—build systems that explain what happened.