Agent application development · Python, LangGraph & TypeScript · 叶立恒
Building AI applications and submitting patches to Tencent/WeKnora and HKUDS/LightRAG · Open to Agent / LLM application development opportunities
I build agent workflows with explicit state, traceable evidence, and clear failure handling.
- Agent workflows — interview planning, question generation, validation and repair, with a separate workflow for answer evaluation. Built with LangGraph and structured intermediate state. (project · write-up)
- Code retrieval & tools — combine keyword and vector search with symbol graphs, graph expansion and call-path tools; expose citations and execution traces in the UI. (project · write-up)
- Integration reliability — investigate HTTP failures, incomplete resource discovery and inconsistent UI state; turn reproducible bugs into small patches and regression tests. (upstream PRs)
- AI Interview Studio — an interview practice application that turns a resume and job description into tailored questions, supports regeneration, and evaluates written answers. LangGraph + FastAPI + PostgreSQL + Next.js.
- CodeRag — an Agentic GraphRAG application for code repositories, with symbol-aware retrieval, call-path exploration and evidence-backed answers. Tree-sitter + KuzuDB + Chroma + LangGraph.
Selected submitted patches — all four PRs are open, not yet merged, as of September 23, 2026. Links below show their latest status.
- LightRAG: model API errors — reject unsuccessful HTTP responses before consuming generated text, streams or embeddings; add regression coverage and fix a CI mock compatibility issue. (#4057 · write-up)
- WeKnora: resource discovery — follow GitLab pagination headers so project discovery returns more than the first 100 resources; test later-page failures. (#3624 · write-up)
- WeKnora: multimodal inspection — add OCR and image-description chunk filters, preserving the selected type across reloads and resetting it when returning to full-text view. (#3548 · write-up)
- WeKnora: developer docs — clarify Swagger access under Docker Compose, including backend ports, runtime mode and container recreation. (#3620 · write-up)
Reproduction steps, validation scope and AI-assisted development details are documented in the linked PRs and write-ups.