Privacy-First On-Device AI Home Hub

Design "[INSERT PRODUCT NAME]", a privacy-first local AI home hub orchestrating devices from [INSERT ECOSYSTEM, e.g., Matter and HomeKit plus legacy IR appliances]. Present the architecture first; execute after approval. 1. MISSION: keep voice, camera, and household data local by default; the cloud is justified per feature, never assumed. 2. HARDWARE: NPU TOPS target, a mic array, physical mic and camera switches, a secure element, and local redundant storage. 3. LOCAL STACK: wake word, ASR, intent LLM, TTS, and vision all on-device, with specified quantization and latency budgets. 4. DEVICE CONTROL: Matter, Thread, Zigbee, BLE, and IR bridges with local automations that keep working when the WAN is down. 5. AGENT SKILLS: scene orchestration, energy optimization, presence rules for elders and children, and temporary guest access. 6. DATA FLOW: sensor bus to a local event bus to local models; outbound domains are deny-listed and the data-flow map is user-visible. 7. VOICE UX: natural multi-turn commands, follow-ups, household voice profiles, and clarifying questions. 8. SECURITY: secure boot, signed updates with rollback, camera network segmentation, mTLS, and a breach-response runbook. 9. IDENTITY: least-privilege per-member profiles, child restrictions, and no cross-household learning without consent. 10. TRANSPARENCY: a local dashboard showing every listening or vision event, how it was processed, and what left the home. 11. COMPLIANCE: GDPR-style right-to-explanation for automated decisions, data deletion, and audit-ready logs. 12. INTEROP: an open API for self-hosters, Home Assistant-style integration, and export in standard formats.