AI TEACH Innovators: A Privacy-First Conversational Business Intelligence Platform using Hybrid RAG, Knowledge Graphs, and Agentic AI with Local LLM Inference

Modern enterprises face growing complexity in extracting actionable intelligence from distributed, heterogeneous data sources while simultaneously facing regulatory pressure to maintain data privacy and sovereignty. Traditional cloud-based AI solutions require transmitting sensitive business data to external servers, making them fundamentally incompatible with GDPR, HIPAA, and sector-specific compliance frameworks. This paper presents AI TEACH Innovators, a production-grade, fully offline conversational business intelligence (BI) platform that integrates Hybrid Retrieval-Augmented Generation (RAG), Knowledge Graphs, and a three-node Agentic AI pipeline to enable natural-language querying of enterprise data entirely on-device. The system leverages Ollama as a fully local Large Language Model (LLM) inference engine routed through a LiteLLM proxy, guaranteeing that no data leaves the user's premises at any stage. A five-layer guardrail engine enforces input safety, document scoping, semantic relevance filtering, Personally Identifiable Information (PII) redaction, and hallucination detection. Knowledge Graphs powered by Neo4j provide causal reasoning over Key Performance Indicator (KPI) relationships, while ChromaDB and sentence-transformer embeddings enable dense semantic retrieval. Evaluation results demonstrate sub-15-second end-to-end response times, 100% offline operation, and robust hallucination mitigation with an 87.5% detection rate in the self-critique layer.