AI Agent Engineer · RAG & Multimodal Document Understanding · CS Student @ Shanghai University
I'm an undergraduate student majoring in Computer Science and Technology at Shanghai University, ranking in the top 3% of my major.
My current interests focus on:
- AI Agents and long-term memory systems
- RAG systems with hybrid retrieval, reranking, and multi-hop reasoning
- Multimodal document understanding, OCR, and layout analysis
- LLM application engineering with FastAPI, vector databases, and evaluation pipelines
- Model adaptation and deployment across PyTorch, MindSpore, and Transformers ecosystems
I enjoy building AI systems that are not only experimental, but also usable, maintainable, and deployable in real-world scenarios.
- 🤖 AI Agent systems
- 🧠 Long-term memory for LLM applications
- 🔎 Retrieval-Augmented Generation, Hybrid Retrieval, Rerank
- 📄 OCR, document parsing, and multimodal understanding
- 🧩 Model migration, training, fine-tuning, and evaluation
- ⚙️ Backend engineering for AI applications
A long-term memory system for project management scenarios, built with FastAPI, Feishu Open Platform, SQLAlchemy, SQLite, and ChromaDB.
The system captures discussions from Feishu group chats and document comments, extracts project decisions, reasons, objections, owners, deadlines, and project stages, then actively retrieves historical decisions in later conversations.
Highlights
- Built Feishu event callback interfaces with FastAPI
- Designed structured memory schemas for project decisions
- Stored long-term memory in relational tables and vector databases
- Implemented semantic retrieval, keyword fallback, and time-aware ranking
- Supported decision conflict detection and memory update mechanisms
- Added automated tests for event callbacks, memory storage, retrieval, and conflict updates
A command-line memory system designed for developers, supporting command recording, workflow reuse, natural language retrieval, and context-aware suggestions.
Highlights
- Implemented commands such as
mem memorize,mem search,mem watch,mem workflow, andmem suggest - Designed metadata for directory, command pattern, execution feedback, and usage frequency
- Combined vector retrieval, keyword fallback, and CLI-specific ranking
- Supported contradiction handling for outdated or corrected memories
- Evaluated retrieval quality using Hit@1, Hit@3, MRR, and operation-saving metrics
A multi-stage RAG system based on FAISS and large language models, integrating hybrid retrieval, reranking, active retrieval, multi-hop retrieval, and memory gating.
Highlights
- Combined dense retrieval with keyword retrieval using RRF
- Added reranking to improve context relevance
- Implemented active retrieval for dynamic query expansion
- Supported multi-hop reasoning for complex questions
- Designed memory gating with importance scoring, time decay, and semantic deduplication
A multimodal document parsing system that converts images and PDFs into structured Markdown / JSON outputs.
Highlights
- Integrated layout analysis, formula detection, OCR, and post-processing models
- Used models such as doclayout_yolo, YOLOFD, UniMERNet, and PaddleOCR
- Designed spatial alignment strategies for text, formulas, and layout blocks
- Built a unified JSON schema and JSON-to-Markdown conversion pipeline
- Optimized multi-model inference workflow for stability and extensibility
Shanghai Shoukou Technology Co., Ltd.
Nov 2024 - Sep 2025
- Built data generation pipelines for OCR and document understanding tasks
- Generated over 10M+ high-quality training samples
- Optimized training strategies and data distribution using Swift training framework
- Developed backend services with FastAPI
- Designed automated evaluation pipelines for multi-model comparison
- Improved data quality control for annotation and review workflows
Institute of Software, Chinese Academy of Sciences
Jun 2024 - Nov 2024
- Contributed to the MindSpore / MindNLP ecosystem
- Migrated HuggingFace Transformers models to MindSpore
- Implemented Data2VecVision adaptation in MindNLP Transformers
- Optimized training on Ascend hardware
- Built a text-driven image segmentation system based on CLIPSeg
- Solved issues related to model structure differences, operator compatibility, and weight conversion
- 🥇 First Prize, China International College Students' Innovation Competition, Shanghai University Division, 2025
- 🥈 Provincial Second Prize, Lanqiao Cup National Software and IT Competition, 2025
- 🥈 Silver Award, iFLYTEK Cup Programming League, 2024
- 🥉 Third Prize, Huawei Zhilian Cup Wireless Programming Competition, 2024
- Email: lmh04102@163.com
- GitHub: https://github.com/LuMH1027
I'm currently interested in long-term internship opportunities related to AI Agents, RAG, multimodal document understanding, and LLM application engineering.