Zook Team

A Context-Aware Personal AI Assistant on Smart Devices: Protecting Users and Supporting Everyday Life SiAssistant is a context-aware personal AI assistant built cross-platform around one shared Rust core: the macOS app is fully supported today and Android is in progress. With the user's permission, it reads the current screen content and serves two jobs at once: everyday help and realtime protection. For everyday life, one hotkey summarizes long articles, translates paragraphs, solves questions step by step, or analyzes selected content in floating panels. For protection, an offline rules engine scores every screen snapshot 0-100 in milliseconds at zero token cost: keyword groups, impersonation and pressure patterns, plus extracted entities (links, phone numbers, bank accounts) and dangerous combinations. Scores 70 and above trigger a red alert; 35-69 mark content suspicious. A built-in RAG layer lets previously confirmed cases vote on new text, so paraphrased scams are caught even with no new keyword match, and an auto-learn loop distills reviewed cases into weighted rules with snapshots and undo. Dangerous hits are cross-checked by AI with a fixed verdict format: Summary, Verdict, Reasons, Action. One shared Rust core (via UniFFI) owns all logic — scoring, PII redaction, rule moderation, prompts, and AI calls — while Swift/Kotlin handle platform I/O only. Privacy is structural: sensitive apps are never read, PII is redacted in two layers before anything leaves the device, screenshots stay in RAM, and API keys live only in Keychain/Keystore plus RAM.