Many organizations struggle with information locked in documents, making it difficult for employees to find specific details efficiently. Traditional manual searches are time-consuming, and commercial document AI products often have opaque pricing and data residency concerns. To address this, the author developed AI-DocumentIntelligence, an open-source, self-hostable, and provider-agnostic RAG platform. This platform aims to be transparent, swappable, and auditable for handling sensitive documents. The core functionality involves ingesting documents, splitting them into meaningful chunks, embedding these chunks into PostgreSQL using pgvector, and answering natural-language questions. A key feature is the ability to switch between LLM providers like OpenAI and Anthropic via a simple configuration. The tech stack is intentionally standard, using React, Node.js, LangChain, and PostgreSQL, making it easy to deploy for teams already familiar with these technologies. The architecture separates the UI, API, document processing, and AI layers, with retrieval and generation processes being distinct for easier debugging. Key learnings emphasize the importance of provider abstraction, well-tuned chunking, and robust local development setups. Developing with credential-gated APIs presented challenges, leading to validation with local embedding models. The author views this project as part of a larger effort to apply LLM orchestration to real-world workflow problems encountered in digital service delivery. The project is available on GitHub for local development and contributions, with detailed setup instructions in the README.