Will Gao

I architected an Agentic AI assistant that automatically creates rich lesson content using a retrieval-augmented generation (RAG) back end. The system ingests structured course outlines, lecture transcripts, and subject-matter PDFs, which I chunk and embed into a semantic store.

Using LangChain and LlamaIndex as the core RAG technologies, the platform retrieves relevant knowledge for each learning module and drives text generation pipelines. I implemented FastAPI to serve the RAG endpoints, while Amazon Bedrock hosts the underlying large language models.

To maintain a consistent tone and pedagogical style, I designed custom prompt engineering logic. Additionally, I built analytics pipelines to measure content coherence and track student engagement signals. Through iterative improvements, I increased coherence scores by 35%.

I architected an AI-powered document search, summarization, and template-generation tool tailored for an industrial report editing platform. The system uses a hierarchical RAG approach that retrieves relevant report sections at multiple levels—document, chapter, and chunk—by leveraging metadata-aware indexing and summary embeddings.

Users can select multiple reports, and the system iteratively aggregates summaries across similar chunks. This enables AI to generate cohesive templates across sections, capturing common patterns and producing structured outputs.

I built the ingestion pipeline using PDF parsing combined with optical character recognition (OCR) and chunking and vector indexing in OpenSearch. Leveraging LangGraph and LangChain, I implemented multitiered retrievers and a self-reflective agentic flow. FastAPI serves the search, summary, and template-generation endpoints.

This modular architecture enables users to generate high-quality draft templates from five to ten reports in under three minutes.

I designed and implemented an end-to-end MLOps system to support the deployment of machine learning workflows in production for a multi-region cloud environment. The architecture combined Amazon SageMaker Pipelines with Apache Airflow (Amazon MWAA) for orchestration, enabling automated training, validation, and deployment cycles.

To ensure reliability and reproducibility, I provisioned infrastructure using Terraform and integrated monitoring and CI/CD hooks to support model lifecycle management. The pipeline was designed to scale with minimal manual intervention, enabling rapid iteration while maintaining high observability and control.

This system powers clients' NLP and classification use cases, allowing clients to iterate on models quickly, reducing time-to-production from days to minutes.