I work at the intersection of Software Engineering (SE) and AI, with a focus on making LLM-assisted software development more reliable, reproducible, and scalable—especially for cross-file code changes and repository-level workflows.
Right now, I’m particularly interested in structured “agent instruction” documentation (e.g., CLAUDE.md / AGENTS.md / project rules) and how SE techniques like ADR, C4/arc42 architecture documentation, and interface/contract specs can strengthen these artifacts to improve correctness and reduce rework.
- 🔭 Current focus: Empirical research on LLM coding assistants for cross-file modifications (effectiveness, generalisability, failure modes)
- 🧪 Methods & interests: Empirical SE, repo mining, benchmark-driven evaluation (e.g., Repo-level tasks), failure analysis, tooling/protocol design
- 🧰 Also exploring: RAG reliability (evaluation, retrieval/grounding, guardrails) and practical agent pipelines
- 💬 Ask me about SE × LLMs, repo-level evaluation, structured documentation protocols, or building reliable LLM tooling
- 📫 Reach me: PeixuanXia@gmail.com
- 😄 Pronouns: She/Her
- ⚡ Fun fact: I love traveling—especially places with great local food and culture
If you're interested in collaborating on AI for SE / SE for AI, feel free to reach out!