🎓 M.S. student, School of Data Science & Engineering, East China Normal University
🔬 Research interests: NLP · LLM hallucination mitigation · Agent engineering
An open-source MCP server that lets LLM agents directly automate the Windows desktop. As a contributor, I got my fix merged the same day the issue was filed (PR #585, closes #583):
- 📦 Large output no longer vanishes — commands printing hundreds of KB lost their entire result to MCP client token limits; the response now keeps the head and the tail with an actionable truncation marker
- ⏱ Timeouts are distinguishable — a timeout no longer shares its exit code with real failures, and the output produced before it ended is surfaced so agents can pick the right retry strategy
- ✅ Layered design + 11 regression tests — capping lives at the tool layer, leaving 10+ internal structured-output call sites untouched; full suite green
- 🥉 Ali Tianchi × CCKS 2026 — Multimodal Knowledge Graph Construction for Cultural Tourism — National Third Prize · competition page
- Kaggle × NVIDIA Nemotron Model Reasoning Challenge — Top 20% (831 / 4185, Mar – Jun 2026) · competition
LLM training: SFT / LoRA fine-tuning · RLHF (GRPO / PPO) · multi-GPU training & inference · on-policy distillation (OPD)
Application engineering: agent development · MCP / skill design · Context Engineering
