To optimize the economic viability of LiveAssist, an "always-listening" meeting copilot, engineers implemented a multi-stage filtering architecture to address massive token over-consumption caused by continuous audio streams. In typical 60-minute meetings, approximately 70% of dialogue consists of non-actionable small talk, which previously led to excessive API costs and high latency when piped directly into RAG pipelines. The team replaced naive full-stream ingestion with a three-tier system: first, a sub-50ms lightweight intent classifier acts as a fast semantic router to drop casual banter; second, a context condenser performs query normalization and retrieves relevant chunks from an AES-256 encrypted corporate vector database; and third, the system generates concise answers for the host UI. By decoupling speech ingestion from the core RAG generation through this upstream gating, the developers slashed input and output token usage by over 65% for 45–60 minute calls. Furthermore, this architectural shift reduced response latency to near real-time by eliminating false-positive RAG lookups and ensured that the host UI remains free of conversational noise.