I am an M.S. candidate in Communication Engineering, working on causal structure learning and generative models.

My current engineering focus is AI Agent and LLM application development, especially context engineering, tool execution, workflow orchestration, long-running task reliability, and observable AI systems.

• Building reliable backends for AI Agent applications
• Context engineering and tool-based agent workflows
• Persistent state, approval flows, recovery, and result validation
• RAG systems with document lifecycle and retrieval observability
• Causal structure learning for generative and visual models

An experiment-driven Research Agent Runtime for computational research workflows.

The system is designed around codebase understanding, experiment planning and approval, persistent workflow state, long-running execution, failure recovery, and result validation.

Current progress includes the overall architecture, V1 scope, core domain model, versioned research context, controlled tool execution, and reliable task orchestration design.

Python asyncio FastAPI SQLAlchemy 2.0 PostgreSQL Alembic
Context Engineering Tool Calling Workflow Orchestration Agent Runtime

A RAG-based ticketing Agent engineering prototype for enterprise IT support.

It connects document ingestion, manual indexing, retrieval, ticket classification, approval-based ticket creation, and Agent execution auditing into an end-to-end workflow.

Key features include:

• Document upload, indexing, retrieval, and deletion lifecycle
• RAG answers with source attribution and low-relevance rejection
• Knowledge search, ticket classification, and ticket creation tools
• Preview-confirm execution flow for controlled write operations
• Approval-state and server-side draft consistency validation
• Structured logs for Agent runs, tool calls, approvals, and retrieval
• Automated tests, GitHub Actions, Alembic, and Docker Compose

Python FastAPI SQLAlchemy SQLite ChromaDB
RAG Tool Calling Agent Observability Docker Compose

A local-first Windows desktop app for understanding computer usage, work-rest rhythm, reminders, and focus signals.

It automatically tracks foreground applications, applies local classification and privacy rules, stores usage data in SQLite, and provides daily insights, idle review, reminders, and a lightweight desktop companion.

Tauri Vue 3 TypeScript Rust SQLite Local-first

A local Windows tool for restoring Codex Desktop conversations that still exist on disk but no longer appear in the application.

It supports session discovery, metadata repair, automatic backups, rollback, and a GUI workflow for Windows users.

PowerShell WinForms SQLite JSONL Local-first Developer Tools

A reusable CLI tool and AI Skill for extracting, validating, transcribing, and organizing Bilibili video content.

It prefers official subtitles when available, detects suspicious or missing subtitles, can fall back to Whisper after confirmation, and outputs transcript text plus structured metadata for downstream AI summarization and analysis.

Python AI Skill Bilibili API Whisper ffmpeg Structured Output

• Backend: Python, FastAPI, asyncio, SQLAlchemy, PostgreSQL, SQLite
• Agent Systems: Context Engineering, Tool / Function Calling, controlled execution, state and workflow orchestration
• LLM Applications: RAG, document lifecycle, source attribution, retrieval logging
• Desktop & Product Engineering: Tauri, Vue 3, TypeScript, Rust, SQLite
• Engineering: Pytest, Alembic, GitHub Actions, Docker Compose, Ruff, mypy
• Research: causal discovery, causal representation learning, generative models, counterfactual image generation

My research focuses on causal structure learning for diffusion-based generative models, including causal discovery, causal representation learning, and counterfactual image generation.

I have submitted a first-author manuscript to Neurocomputing.

• GitHub: Air000000
• Email: 1113476369@qq.com