What problem does it solve? Building real-time voice AI platforms requires solving hard distributed-systems problems: sub-second voice turn latency, stateless token verification across microservices, stateful conversational turn management, and grounded regulatory answers. This Skill provides a complete architectural blueprint with reference guides and copy-pasteable code patterns extracted from a production ATC voice simulator platform. ## Core Features & Use Cases - 7-Layer Redis Caching Engine: Reduce voice pipeline latency from ~2,600ms to under 280ms using layered caches for embeddings, RAG grounding, session state, JWKS keys, rate limiting, and TTS audio. - Zero-Trust RS256/JWKS Authentication: Implement asymmetric JWT signing, stateless public-key verification with key rotation, and rotating refresh token families with replay-attack detection. - LangGraph Voice State Machine: Orchestrate half-duplex push-to-talk turns using interrupt boundaries, additive transcript reducers, and fuzzy slot validation. - Use Case: When architecting a new voice-driven AI service, apply the microservice monorepo pattern (server.js vs app/app.js separation, per-service MongoDB isolation, K8s probes) and the 4-layer React frontend pattern (Components → Hooks → Services → Redux slices) to structure the codebase. ## Quick Start Ask the AI to design a low-latency voice agent backend using the 7-layer Redis caching pattern and LangGraph interrupt boundaries from this architecture skill.
