A live voice-agent prototype exploring how private AI, operational guardrails, tool use, memory, analytics, and human escalation can work together in a modern banking experience.
Qwen 14B, MLX Whisper Large V3 Turbo, Silero and Kokoro run in the private local inference environment.
Meaningful turns, tools, escalation, failures and responses are auditable through n8n.
Prompt-injection defense, synthetic data, human escalation and audit controls protect the demo boundary.