AI Agent Engineer / AI Application Researcher
Agent Runtime · LLM Workflows · AI Product Engineering · Game AI · Computer Vision
I build AI systems that turn messy context into usable tools: agent workflows, memory systems, content intelligence, and decision-making environments.
- Building agentic applications across content creation, knowledge management, virtual characters, and game-like decision systems.
- Interested in LLM workflow engineering: tool use, long-running jobs, memory, evaluation loops, streaming UX, and deployment.
- Experienced in game AI and simulation: reinforcement learning, MCTS, strategy environments, Unity/UE workflows, and virtual character systems.
- Previously worked on computer vision research, especially UAV tiny-object tracking, datasets, benchmarks, and reproducible experiments.
| Direction | What I build | What it shows |
|---|---|---|
| Agent Engineering | Multi-step LLM workflows, tool orchestration, long-running task pipelines, SSE progress streams | Turning agents from demos into usable software |
| AI Application Research | Content intelligence, AI analysis services, cross-platform data products, personal knowledge systems | Mapping business/product questions into model-assisted systems |
| Autonomous Agents | Persona agents, memory layers, heartbeat loops, intent/state engines, sub-skill routing | Designing agents that maintain context and act over time |
| Game & Strategy AI | RL/MCTS agents, game simulators, procedural content generation, virtual character interaction | Handling complex state, uncertain actions, and evaluation |
| Research Systems | Dataset design, tracking benchmarks, reproducible evaluation pipelines | Making experiments measurable and comparable |
| Project | What it demonstrates |
|---|---|
| AI Script & Storyboard Agent | Full-stack agent product for live/script/storyboard generation with Next.js, Hono/FastAPI, PostgreSQL, Docker Compose, and streaming progress UX. |
| Autonomous Persona Agent Runtime | Long-term memory, emotion/state engines, heartbeat loops, intent pools, content-source adapters, and pluggable agent skills. |
| Cross-platform AI Knowledge Manager | Browser extension + web app + shared adapters for collecting, syncing, searching, tagging, and summarizing information across platforms. |
| MetaGPT Minecraft Contributions | Multi-agent open-source contributions: action developer, curriculum agent, skill manager, Mineflayer integration, and event loop fixes. |
| xiaohongshu-skills / douyin-skills | CDP-based platform automation skills, account isolation, browser process control, human-like interaction, publishing/search/ops workflows. |
| OLOD | UAV tiny-object tracking dataset and benchmark; research execution from data construction to evaluation protocol. |
| Direction | Tools & Skills |
|---|---|
| AI / Agents | Python, TypeScript, LangGraph, Agno, OpenAI-compatible APIs, RAG, prompt/workflow design, agent evaluation |
| Product Engineering | Next.js, React, Hono/FastAPI, PostgreSQL, Redis, Prisma, Supabase, Docker, Nginx, Vercel |
| Automation | Chrome CDP, browser extensions, platform adapters, scheduled jobs, data normalization, anti-fragile scraping workflows |
| ML / Research | PyTorch, OpenCV, RL, MCTS, dataset design, benchmark evaluation, experiment tracking |
| Game / Interactive | Unity, Unreal Engine, C++, C#, procedural generation, virtual character systems |
- Making agents useful after the first prompt: memory, state, schedules, tools, and recovery.
- Designing AI products where the model is only one part of a larger workflow.
- Building virtual characters and game worlds that can reason, adapt, and generate stories through interaction.
- Studying open-source agent frameworks and extracting practical architecture patterns.
- Blog: halyu.me
- GitHub: github.com/yuymf
- Email: yuymf@qq.com