AI Algorithm Engineer based in Shanghai, focused on LLM applications & Agent systems.
7 years of engineering experience — 5 years building large-scale data infrastructure at NIO (autonomous driving data) and SAIC-GM (intelligent connected vehicles), and 2+ years transitioning into LLM application development. Currently building a multi-agent Q&A platform for the banking domain.
I work at the intersection of data engineering and LLMs — turning messy enterprise data into reliable AI products.
- Bank Intelligent Q&A Platform — Multi-agent architecture with NL2SQL, RAG, and tool-calling routing. Built on LangGraph with BGE-M3 hybrid retrieval and Qwen-series models.
- Medical RAG System — Tri-modal hybrid retrieval (dense + sparse + ColBERT) over Milvus/HNSW, with LangGraph orchestration and RAGAS evaluation. ChatGLM3-6B fine-tuned with LoRA/QLoRA.
- Exploring — Multimodal agents, agent evaluation frameworks, and production deployment patterns (vLLM, quantization).
LLM & Agents
LangGraph · LangChain · RAG · NL2SQL · vLLM · LoRA/QLoRA · RAGAS
Models & Frameworks
Qwen · ChatGLM · BGE-M3 · PyTorch · Transformers · PEFT
Data & Infra
Milvus · Doris · ClickHouse · Hive · Iceberg · Spark · Flink
Languages & Tools
Python · SQL · Java · Docker · Git · Linux
| Project | Stack | Highlights |
|---|---|---|
| Medical RAG System | FastAPI, LangChain, Milvus, Qwen, ChromaDB | Retrieval-first multi-turn clarification dialogue for medical education & clinical support, SufficiencyJudgement state machine for dynamic query refinement, medical knowledge base chunking with hybrid BM25+dense retrieval, support for clinical decision-making queries with cited evidence, end-to-end integration tests with state transition coverage |
| Smart Crib Guard | PyTorch, Qwen2.5-VL-7B, PEFT (LoRA), Transformers, OpenCV | 9-class baby behavior recognition via Qwen2.5-VL-7B LoRA fine-tuning (Loss: 0.0175), automated video frame extraction + multi-strategy augmentation (brightness/rotation/crop/flip), JSON-standardized annotation pipeline, real-time inference with confusion matrix generation, edge-cloud integrated monitoring |
| Legal RAG System | FastAPI, ChromaDB, bge-zh, Qwen3 | Statute-aware article-level chunking, hybrid BM25+dense retrieval, query rewriting, cited answers with confidence scoring |
| Positive Chinese Chatbot | PyTorch, ChatGLM3-6B, LoRA, Gradio | LoRA fine-tuning on Douban "夸夸" corpus for encouraging replies, Trie-based dirty-word filter with variant/homophone detection, multi-strategy decoding with BLEU/ROUGE/diversity evaluation |
| Chinese Sentiment Classifier | PyTorch, BERT-base-Chinese, Transformers, ChnSentiCorp | Selective last-N layer unfreezing on BERT with weighted sampling and jieba synonym augmentation, AMP + warmup-linear LR + early stopping, MC-dropout uncertainty and attention-based explainability, 94.5% Acc / AUC 0.983 |
Open to opportunities in LLM / Agent engineering — particularly roles involving data agents, RAG systems, or multimodal applications.
Feel free to reach out via GitHub.