AI Application Developer — building production-grade AI Agent systems from scratch.
10万+ task requests handled in real production systems. 161 tests, 0 failures. 14/14 penetration tests passed. Deployed 24/7 on cloud.
I build autonomous AI agent systems that don't just call an API — they improve themselves, stay secure, and run reliably in production.
A complete autonomous agent framework with 9 self-improving engines. Built from scratch over one semester.
Key numbers:
- 161 tests, 0 failures
- 14/14 penetration tests passed (OWASP LLM Top 10)
- 1000/1000 stress test requests, P95=150ms
- 9 autonomous engines: self-evolution, debate, bootstrap, meta-agent
- Deployed 24/7 on Alibaba Cloud ECS
What it does: The system doesn't just answer questions — it detects its own capability gaps, generates missing tools at runtime, evolves tool implementations, and rolls back on regression. All 9 engines have real LLM-verified code.
→ Live Demo | GitHub | Blog (CN)
Research on how reasoning strategies transfer across modalities in multi-modal AI models. Found that visual-to-text strategy contamination is significantly stronger than text-to-visual (p = 0.008).
Why it matters: If you're building GPT-4V / Gemini-style multi-modal systems, your visual training data may be silently degrading text reasoning performance.
Paper: arXiv preprint (in submission) | GitHub
| Category | Technologies |
|---|---|
| AI/LLM | DeepSeek V4, Qwen2.5, Claude (API), OpenAI-compatible APIs |
| Frameworks | FastAPI, LangChain, Streamlit, AsyncIO |
| Agent Tech | RAG (ChromaDB + all-MiniLM), Tool calling, Multi-agent orchestration |
| Security | OWASP LLM Top 10 audit, penetration testing, sandbox isolation |
| Deployment | Docker, Alibaba Cloud ECS, systemd, Prometheus |
| Languages | Python (primary), Java / Spring Boot, SQL |
| Project | What It Does | Stack |
|---|---|---|
| ai-agent-playground | 9-engine autonomous agent, self-evolving, 24/7 production | Python, FastAPI, DeepSeek V4, ChromaDB |
| aettl-research | Multi-modal AI research, NeurIPS 2026 submission | Python, NumPy, SciPy, LaTeX |
| ling-reader (灵阅) | Multi-Agent video production system | Agent orchestration, video rendering |
| ling-producer (灵映) | Agent video rendering pipeline | Agent system design |
- Looking for: AI Application Developer / AI Agent Engineer roles
- Location: China (open to relocate)
- Available: Immediately
- 从学生项目到生产级 AI Agent — how I built a production AI agent system from scratch
- English version
Market rewards shipping, not learning. — Liu Zewen