Zhichao Zhao

AI Application Developer · Agent Systems Engineer · LLM/RAG Builder

Email: 220246355@seu.edu.cn

Phone: +86 188xxxx7217

---

About

AI-focused software engineer and graduate researcher with hands-on experience in multi-agent systems, retrieval-augmented generation (RAG), LLM alignment, and domain-specific intelligent applications. My work centers on building AI systems that are not only capable, but also explainable, controllable, and production-oriented.

Currently pursuing an M.S. in Software Engineering at Southeast University after recommendation-based admission. I am especially interested in AI application development, agent systems, LLM engineering, and intelligent product prototyping.

---

Education

Southeast University

M.S. in Software Engineering

Sep 2024 – Jun 2027

- Admitted through recommendation-based graduate admission

- Relevant coursework: Algorithms Design and Analysis, Pattern Recognition, Digital Image Processing

China University of Mining and Technology

B.Eng. in Data Science and Big Data Technology

Sep 2020 – Jun 2024

- Relevant coursework: Deep Learning, Machine Learning, Recommender Systems, Big Data Storage and Management

---

Honors

- Co-author of an SCI-indexed paper

- International Second Prize, Mathematical Contest in Modeling (MCM/ICM)

- Provincial Second Prize, “Black Tech” Special Competition, 18th Challenge Cup National Student Extracurricular Academic and Technology Competition

---

Projects

Intelligent Travel Planning Multi-Agent System

Tech Stack: LangGraph, LangChain, Streamlit, RAG, ChromaDB, MCP, SQLite

Built a production-style multi-agent travel planning system for complex user requests, tool orchestration, and personalized multi-turn interaction.

- Architected an explainable multi-agent workflow with five coordinated roles: Main Agent, Planner, Executor, Summarizer, and Feedback Agent

- Designed a dual execution framework with Simple ReAct for lightweight queries and Plan-then-Execute for multi-city, constraint-heavy scenarios

- Combined Qwen for autonomous tool usage with DeepSeek-R1 for structured planning and execution

- Built a hybrid RAG plus MCP architecture integrating local knowledge retrieval with real-time external tools

- Improved Hit@5 from 70% to 84% on a self-built travel evaluation dataset using DashScope Embeddings, ChromaDB, and Cross-Encoder reranking

- Implemented TXT/PDF ingestion, chunking, persistence, similarity retrieval, and MD5-based file deduplication

- Designed a four-layer memory framework covering short-term state memory, long-term user preference memory, persistent session history, and context compression

- Reduced average token consumption in multi-turn conversations by 30%

Intelligent Medical QA and Assisted Triage System

Tech Stack: DeepSeek-R1-Distill-Qwen-1.5B, LLaMA-Factory, LoRA, DPO, SearXNG, ChromaDB

Developed a domain-specific medical assistant system focused on safe generation, retrieval grounding, and preference alignment.

- Designed a full data cleaning pipeline for the Huatuo-26M corpus using regex filtering, heuristic rules, and perplexity-based quality screening

- Curated 100,000 high-quality multi-turn medical QA samples for supervised fine-tuning

- Distilled 2,000 pairwise preference samples and defined an annotation framework favoring clarity, structured output, and medical disclaimers

- Conducted two-stage fine-tuning with LoRA-based SFT and DPO alignment on a DeepSeek-1.5B backbone

- Achieved a 90% safe refusal rate in high-risk diagnosis and treatment-inducing scenarios

- Built a Web-RAG pipeline using AnythingLLM, bge-large-zh, ChromaDB, and SearXNG to integrate up-to-date medical references and reduce hallucination

---

Research

Heart Rate Prediction from Millimeter-Wave Radar Vital Sign Signals

First Author

Manuscript pending acceptance at Biomedical Signal Processing and Control (CAS Q2)

Tech Stack: PyTorch, Temporal Attention, Bi-LSTM, TCN, Millimeter-Wave Radar Signal Processing

- Designed ERNet, a deep spatiotemporal architecture for robust heart rate prediction under noisy sequential conditions

- Combined TimeDistributed CNN for spatial representation learning with Bi-LSTM for long-range temporal dependency modeling

- Introduced a temporal attention mechanism to suppress abrupt local artifacts and improve robustness

- Proposed a PCA-based adaptive fusion strategy to address multi-channel imbalance

- Built a complete end-to-end PyTorch pipeline for training, inference, and evaluation

- Achieved 97.6% prediction accuracy, outperforming baseline methods

---

Technical Skills

Programming: Python, C++

Frameworks and Tools: PyTorch, LangChain, LangGraph, ChromaDB, LLaMA-Factory, Streamlit

LLM Engineering: RAG, LoRA, DPO, MCP, multi-agent orchestration, retrieval systems, preference alignment, context compression

Systems: Linux, MySQL, SQLite

English: CET-6

---

Career Interests

- AI Application Development

- Agent System Development

- LLM Engineering

- Intelligent Product Prototyping

- RAG and Tool-Augmented AI Systems

---

Contact

Email: 220246355@seu.edu.cn

Phone: +86 188xxxx7217