Ziyang Liu

• —VM migration optimization: improved migration workflow (orchestration, data transfer, iterative/incremental); refactored incremental stage with multi-threading to increase throughput and reduce downtime.
• —Virtio injection compatibility: designed VMware→KVM Virtio auto-injection for MBR/UEFI guests; improved initramfs dependency handling and rollback for legacy kernels (e.g., RHEL5 2.6.18), including a safe IDE fallback path.
• —Cross-cluster live migration: developed key features in SMTX OS VM services (live migration, monitoring, alerting); implemented post-migration validation.
• —QGA syscall-trace instrumentation: extended QGA tooling with Windows command invocation and ordered syscall tracing; exported structured logs for faster triage.
• —Kernel debugging: isolated a Linux virtio-serial/console issue where payload >4KB could stall the second poll; produced a minimal reproducer and workaround (chunking) for escalation.
• —Incident troubleshooting: resolved 100+ customer cases; standardized SOP for a UOS kernel-related issue; received formal customer appreciation.

• —Guest tooling (VMTools / QGA): extended VMTools based on QEMU Guest Agent — stats collection, script distribution, fsfreeze orchestration.
• —VM migration tool: implemented a tool to convert VMware VMs into SmartX platform VMs.

• —Built Go + Gin services for recruiting/HR integration and data synchronization with external OA systems.
• —Used MongoDB for sync metadata and retry states; improved APIs with authentication, rate limiting, pagination, documentation, and regression tests.

Research fork of CursorTouch/Windows-MCP. Agent uses the MCP server, patches it, hot-reloads it, and verifies whether real desktop behavior improved. Focused on AI-driven Windows automation reliability.

Local-first vocabulary CLI for word learning. Anki-style cards, FSRS-style spaced repetition scheduler, LLM-powered teaching, and swappable AI companion packs. State stored in SQLite.

PyTorch training toolkit, open-sourced from a garbage sorting WeChat Mini Program. Served model inference via Flask; built GitHub Actions CI for Docker image builds.

2D pixel-style multiplayer loot-shoot-extract game (搜打撤). TypeScript monorepo: Vite + Canvas 2D client and Node.js + WebSocket server with shared Zod protocol validation, MessagePack encoding, 20Hz server tick, and 120ms client interpolation.