- 🔭 Now: AI Native Architect @ Bluesmart Digital — Shipping intelligence from prototype to production
- 🧠 Focus: FDE · AI Native Architecture · Agent Workflow Deployment · RAG / GraphRAG
- 📚 Research: Education Agent — SCI Q1 journal under review · 1 ETNCC conference paper
- 🌱 Community: Silver-tier organizer at Datawhale open-source AI community
- 💬 Reach me: threethreeliu@163.com
🚧 My world revolves around Mercury — Mercury At The Center Of My World
Turning LLM capabilities into shippable, iterable, production-grade engineering systems — this is the through-line of everything I do today.
| Track | Core Capabilities | Toolchain / Keywords |
|---|---|---|
| 🧬 Model Training & Fine-tuning |
SFT · LoRA · Parameter-Efficient Fine-tuning (PEFT) RLHF · DPO · PPO · GRPO — alignment & reinforcement learning Model internals & attention mechanism analysis Distributed training, GPU memory & throughput optimization |
PyTorch · DeepSpeed · CUDA · PEFT · Transformers · TensorFlowTRL · trl-lib · vLLM · Megatron-LM→ GPU resources optimized 25% · dev velocity +40% |
| 🤖 Agent Engineering |
Planning-Execution-Feedback orchestration Multi-agent collaboration · role-based task division Tool calling / state machine / context management |
LangGraph · LangChain · AutoGen · ReAct · Function Calling→ Solvely.ai · Operations Hive — HiveMote |
| 🧪 PromptOps & Evaluation |
Automated failure-sample feedback loop (AutoFeedbackLoop) Template perturbation · multi-model adapter layer · fallback strategy Badcase governance & effect alignment |
PromptOps · A/B Eval · LLM-as-Judge · Regression Suite→ Significant stability gains on long-horizon tasks |
| 🔍 RAG / GraphRAG / Retrieval |
Document parsing · knowledge extraction · Knowledge Map Vector retrieval · hybrid retrieval · Reranking Citation tracing · multimodal QA · multi-document conversation |
RAG · GraphRAG · Vector DB · BM25 · Embedding · Schema→ Solvely Long-Document AI Content Engine |
| 🚀 Production Engineering |
Streaming output · explicit caching · context state machine Token Mask · inference cost governance Canary release · auto-scaling · monitoring & alerting |
Streaming · Cache · Docker · K8s · Linux · Aliyun→ TTFT 35.7s → 17.4s (-51%) · Token cost -90% · throughput +91% |
🌌 "Mercury At The Center Of My World" — ⭐ From threethreeliu.top