What are the architectural patterns for building offline-first AI applications?

Offline-first AI architecture ensures applications function without connectivity while synchronizing when online. Core pattern: local model handles all inference, local storage caches data and model artifacts, sync engine reconciles local and cloud state when connectivity is available. Architecture layers: presentation (UI with offline indicators), application logic (local inference pipeline), local model runtime (Ollama, llama.cpp, or ONNX Runtime), local data store (SQLite, ChromaDB for vectors), sync manager (conflict resolution, queue management), and optional cloud services (model updates, enhanced processing). Implementation: store model files locally with version tracking, implement model update mechanism (background download when online, validate checksum, atomic swap), cache embedding indexes on disk for instant startup, and implement request queuing for cloud-dependent operations. Sync strategy: local-first with cloud enhancement (local model processes immediately, cloud model refines when available), conflict resolution using last-write-wins or merge strategies, and offline mutation log replayed on reconnection. Progressive enhancement: provide basic functionality offline (local model), enhanced functionality online (cloud model for complex queries, real-time data access), and transparent switching with user notification. Storage management: implement LRU cache for model files (keep most-used models locally), compress and prune vector indexes periodically, and set storage budgets per application feature. Testing: simulate offline conditions in development, test sync conflict resolution, verify data integrity after extended offline periods, and measure performance degradation in offline mode.