Welcome to My Blog #

🌟 About Me #

I am Zeren Shen, a Machine Learning Engineer with a focus on LLMs, Generative AI, and 3D Motion Analysis. My journey began with a solid academic foundation from the University of Toronto, where I pursued a Master’s in Mechanical and Industrial Engineering with an emphasis on Analytics, and the University of Waterloo, where I earned a Bachelor’s degree in Statistics and Actuarial Science with a minor in Computer Science. Over the years, I have had the privilege of applying these skills to various projects, where I’ve collaborated with talented teams to develop AI-driven solutions aimed at solving complex real-world problems.

• 📍 Location: Toronto, ON, Canada
• 🎓 Education:
  • Master’s: Mechanical & Industrial Engineering (University of Toronto)
  • Bachelor’s: Statistics and Actuarial Science (University of Waterloo)
• 📬 Contact:
  • Email: zeren71415@gmail.com
  • GitHub: Followb1ind1y
  • LinkedIn: Zeren Shen
  • Resume: Download my Resume

🚀 Professional Experience #

Machine Learning Engineer #

@ ThinkGenAI Lab Inc. - Toronto, Canada
May 2023 - Nov 2024

Research Assistant RAG System #

• Developed a RAG system that enables research teams to search, retrieve, and summarize newly published papers.
• Built an end-to-end pipeline for PDF ingestion, chunking, and vector embedding (Pinecone), integrating LangChain-powered ReAct agents for real-time retrieval, semantic search, and LLM-based Q&A.
• Containerized microservices with Docker, deployed a FastAPI backend for concurrent request handling and built a Discord bot with conversation summarization, memory management, and PDF export features.

LLM-Powered Dialogue Toy Robot

• Integrated Whisper-STT, GPT-3.5, and Google TTS for dialogue robot pipeline, achieving 1.2s end-to-end latency.
• Developed a high-performance FastAPI service with async/await to enable parallel voice processing and LLM inference, handling 20+ requests per minute through request batching.
• Reduced response time by 50% through optimized pipelining, leveraging audio chunk streaming and token-level generation for seamless user interaction.

Research Assistant #

@ Laboratory for Extreme Mechanics & Additive Manufacturing - Toronto, Canada
Sep 2021 - May 2022

• Implemented TensorFlow-based U-Net models for real-time X-ray image segmentation, achieving 0.93 MeanIoU in pore detection and reducing manual inspection time by 40% through automated defect-tracking algorithms.
• Optimized bubble detection latency by 30% via model quantization and multi-threaded data pipelines, deploying the system as Docker containers integrated with industrial 3D printers.
• Publication: Zhang, J., Lyu, T., Hua, Y., Shen, Z., et al. Image Segmentation for Defect Analysis in Laser Powder Bed Fusion. Integr Mater Manuf Innov 11, 418–432 (2022).

Autonomous Driving Research Intern #

@ Zhejiang University May 2018 - Aug 2018

• Led critical data collection initiatives from driving environments to support further training of computer vision models, contributing to the research in self-driving vehicle technology.
• Collected and analyzed data from over 50 driving hours, capturing 100,000+ frames of video data and 500,000+ data points for various sensor modalities, including LiDAR and GPS.

🛠️ Projects #

• Medical-LLM-Fine-tuning: Fine-tunes LLaMA-3-8B on PubMedQA with QLoRA, optimized via DeepSpeed and vLLM for efficient, low-latency medical QA. Deployable via Docker for scalable clinical inference.

• AgentRAG: AgentRAG is a powerful AI-driven system that integrates web crawling, scraping, vector search, and intelligent chatbot interactions to process online information.

• State Farm Distracted Driver Detection: Developed advanced image classification models using ensemble methods to identify distracted driving behaviors from dashboard camera images, improving safety insights and insurance applications.

• Semantic Segmentation of Aerial Imagery: Implemented semantic segmentation models to classify every pixel in aerial images into six categories, enabling precise analysis of urban landscapes and infrastructure.

• Face Mask Object Detection: Built real-time object detection models to identify and classify masked and unmasked faces, addressing public health needs during the COVID-19 pandemic with live-stream and video applications.

Explore more on my GitHub!

📚 Key Skills #