Shane Zhang (张欣耕)

Rutherford, NJ

• Built a fast structured data processing pipeline with Python, reducing processing time from 10 seconds to under 1 second per document, enhancing document handling efficiency by 90%.
• Designed and accelerated a robust RESTful API with Python FastAPI for real-time data ingestion and NLP processing into Postgres SQL database, supporting over 1 million API calls per day.
• Deployed solutions at scale with OpenShift and Apache Spark, streamlining CI/CD procedures with zero downtime over 120 days.
• Leveraged advanced Agentic RAG systems and Knowledge Graphs with RDF and LPG formats using Python and LangChain to improve global retrieval accuracy, increasing the F1 score of AI-driven compliance analysis from 0.32 to 0.71.
• Implemented evaluation metrics and visualizations with LangChain and RAGAS to ensure LLM quality, reducing hallucination rate by 30% and error investigation time by 60%.
• Utilized OpenAI GPT, Claude, and Google Gemini models and fine-tuned open-source LLM models to develop customized and scalable Agentic RAG systems handling over 200,000 API calls per day.
• Developed human-centered evaluation frameworks (RLHF) to assess LLM performance in real-world scenarios, continuously ensuring alignment with user intents for over 2,000 active users.
• Deployed scalable AI-driven RAG systems on AWS, using S3, EC2, Glue, Lambda, SageMaker, and Bedrock for efficient data processing and LLM integration pipelines.
• Developed and deployed FastAPI interfaces and gRPC protocols for efficient API integration across cloud providers like Azure and AWS.

New Brunswick, NJ

• Designed, built, and deployed an Agentic RAG system using Python, NumPy, Pandas, JavaScript, SQL, and Chroma vector database, automating the parsing and summarization of over 5,000 research documents.
• Engineered an advanced Agentic RAG system with OpenAI and React framework, reducing onboarding time by 50% and improving research efficiency by 40%.
• Collaborated with a cross-functional team of 15 academic researchers, integrating over 200 feedback points into the NLP system to align the platform with current and future research objectives.
• Conducted monthly workshops on ML usage for 10+ staff members, achieving 80% adoption of ML-enhanced workflow across the organization.
• Utilized AWS SageMaker for model fine-tuning and AWS Bedrock for serving the models in production environments.

San Francisco, CA

• Integrated NLP-driven features into the Node.js backend with Python PyTorch (C++ CUDA), enabling real-time text generation and summarization functionalities, cutting response times by 70%.
• Conducted data pre-processing and data exploration using Python PySpark, NumPy, and Pandas, ensuring high-quality data integration and leading to a 20% improvement in model training results using PyTorch distributed.
• Optimized deep learning models through quantization and pruning techniques with TensorRT, achieving a 40% decrease in inference latency.
• Produced a semantic graph database using Neo4j and Cypher queries to store and query complex relationships, improving data retrieval efficiency by 60%.