• Graduated from Duke University, completed Master of Engineering in Artificial Intelligence and Production Innovation (AIPI). Looking for AI Engineer related roles.
• My interests include LLM, agent system, RAG, machine learning, and edge AI. I am passionate about leveraging AI to solve real-world complicated problems and am always eager to learn and explore new technologies.

• Duke University [Aug 2024 - Dec 2025]
  Master of Engineering in Artificial Intelligence and Product Innovation link

• The University of Hong Kong [Sep 2019 - Jul 2024]
  Bachelor of Science in Actuarial Science & Computer Science link

• Programming Languages: Python (main), HTML, CSS, JavaScript, SQL, Java, C++...
• Data Science: Pandas, NumPy, Matplotlib, Seaborn, WebScraping(bs4)...
• Machine Learning/Deep Learning: PyTorch, Scikit-learn, numpy, pandas...
• Computer Vision(CV): OpenCV, Object Detection, Object Tracking, Object Segmentation, Bar code, QR code, ArUco Marker
• Large Language Model(LLM): LangGraph, Ollama, HuggingFace, n8n, RAG, Pinecone, Unsloth, Parameter-Efficient Fine-Tuning (PEFT) including LoRA (Low-Rank Adaptation), Text-To-Speech(TTS), Speech-To-Text(STT)...
• Cloud Deployment: Docker, tumx, AWS, Streamlit...
• Hardware: RaspberryPi Development, Arduino Development, Edge AI, BreadBoard, Sensors, Motor Programming, Servo Programming...
• Others: LINUX, API, SSH, Git...

• Next generation cross-platform personalized AI agent. Sub project: AI agent on VR Glasses.
• Anti Money Laundry machine learning detection extended from 2025 DTCC AI Hackathon.

• Award: Academic Institution Team Award
• Description: Sentinel AI is an AI-powered risk scoring system designed for blockchain wallets. It aggregates and analyzes over 1 billion transactions, focusing on stablecoins (USDT/USDC), which dominate the digital asset market with $2.90T in monthly transfer volume. The platform assigns real-time risk scores based on transaction history, smart contract associations, and monitored KYC transactions, effectively closing compliance gaps.
  
• Technology highlights:
  • Data sourcing: Google Big Query, SQL...
  • Machine Learning: XGBoost, CatBoost, LightGBM, regression models...

• Award: First Rank in AI Tool Track; Top 5 in Overall
  
• Description: The project created during Duke AI Hackathon 2024. The Flashback system is a personal digital brain that remember everything for human to increase memory storage. Feel free to explore project website on devpost project page.
• Technology highlights:
  • Hardware: Raspberry Pi with camera and microphone for information collection (sensors).
  • Software: Full stack website application for processing, storage and hosting; gemini-1.5-pro large vision language model for handling queries requests. In the future, AI agents for executing different tasks and computer vision for face detection is planning to be added.

• Category: Education, Automation
• This project was a sub-project of HKU 𝐀𝐫𝐚𝐫𝐚𝐭 𝐏𝐥𝐚𝐢𝐧 𝐒𝐨𝐮𝐭𝐡𝐞𝐚𝐬𝐭 𝐀𝐫𝐜𝐡𝐚𝐞𝐨𝐥𝐨𝐠𝐢𝐜𝐚𝐥 𝐏𝐫𝐨𝐣𝐞𝐜𝐭, aiming to streamline and automate the data collection and processing during archeological activities. Detailed information could be found in APSAP repo.

• Dual-Stream Emotion XAI Dual-Stream Emotion XAI (Harrisous/XAI-Final-Project) is an interactive Streamlit XAI app that uses Wav2Vec2-based acoustic emotion recognition, RoBERTa-based semantic sentiment/emotion analysis, and Whisper ASR to compare how you sound (prosody, pitch, speed, energy) versus what you say (textual meaning) in side-by-side 3D PCA latent spaces built from deep embeddings and reference datasets (renumics/emodb and dair-ai/emotion), highlighting clusters, nearest neighbors, and cross-modal conflicts such as sarcasm for explainable, reproducible emotion analysis with Hugging Face, Plotly, scikit-learn, and joblib.​ Interactive Webpage

• Duke Chatbot - Blue Devil Bot.
  A chatbot for Duke related information retrieval, built using LangGraph framework for agentic search. Toolbox includes: API calls for exisitng APIs, Web Scrapping, and Tavily search for general search. WebScraping(bs4), LangGraph, Agentic Search, API are the key technologies used in this project.
• LLM Text-To-SQLs Fintuning.
  A large language model (LLM) fine-tuning task using the Bird database, aimed at enhancing LLM's capabilities in automatic text-to-SQL generation. We achieved 50% accuracy improvement on simple Text-To-SQL generation tasks at the expense of the capability to deal with challenging tasks. Benchmark source: BIRD-SQL: A BIg Bench for Large-Scale Relational Database Grounded Text-to-SQLs. Unsloth, Parameter-Efficient Fine-Tuning (PEFT) including LoRA (Low-Rank Adaptation) were used in the project.

• Group Project - AIPet (Pido).
  Our team built an AI pet robot that can follow the owner's commands with tracking function, providing comfort and companionship to the user. The pet is able to move or follow the user, understand instruction bark, show emotion on LCD screen and wag tail. The whole system runs on a RaspberryPi 4B powered by a powerbank and the action are done by motors (powered by another set of batteries) and servo. Presentation url. Key technologies used: sensor integration, GPIO control, speech recognition, computer vision, motor and servo control.
• Driving behavior detector.
  An individual project on raspberrypi 4B with traditional ML model to detect user's driving behavior via acceleration and gyroscope reading and predict the driver's behavior from a trained model. Presenttaion URL

• Movie Recommendation System This project implements an advanced hybrid recommendation system that combines collaborative filtering, content-based filtering, and deep learning techniques to provide personalized movie recommendations to users. The system leverages both user behavior data and descriptive content to create a robust recommendation engine that addresses cold-start problems and provides explainable recommendations.
• Stock Market Transformer Stock Market Transformer is a deep learning project for stock market prediction using autoencoder-based environment embeddings and Transformer (with Mixture-of-Experts and multi-head attention) to forecast future prices from complex time-series market data. The approach delivered promising short-term opening price prediction results with a working app, though further improvements are needed for broader targets and lower loss.
• Sentiment Analysis A project containing different sentiment analysis methods.

• Email: harryli12321@gmail.com
• LinkedIn: www.linkedin.com/in/haochen-harry-li
• GitHub: https://github.com/Harrisous

Feel free to reach out to me for collaboration or just to say hi! 😊