AI-Powered Meeting Operations Assistant Using Speaker Diarization, RAG, and LLMs

Online and hybrid meetings produce lengthy multi-speaker conversations in which decisions, responsibilities, deadlines, and unresolved questions are difficult to recover.This paper proposes an AI-powered Meeting Operations Assistant combining automatic speech recognition (ASR), speaker diarization, retrieval-augmented generation (RAG), and large language models (LLMs).Audio is transcribed and aligned with anonymous speaker labels and timestamps.Transcript segments are indexed with meeting, time, speaker, and accesscontrol metadata.A hybrid retrieval layer supplies relevant transcript passages and approved project documents for summarization, decision extraction, action-item generation, and question answering.Each generated item links to supporting evidence; missing owners and deadlines remain unresolved rather than guessed.Human review is required before any output is sent to external tools or converted into tasks.The paper presents a modular architecture, traceable data model, grounded-generation procedure, and evaluation protocol using public meeting and summarization datasets.Metrics include word error rate, diarization error rate, action-item precision and recall, evidence-citation accuracy, human-rated factuality, and latency.Because the prototype has not yet been evaluated, no performance figures are claimed.The paper is therefore a testable system blueprint rather than a report of completed experiments.