Appreciate the thoughtful read! That exact gap between "privacy policy promises" and actual system architecture is what drove every decision in Kinetib. On the OCR side, avoiding cloud vision APIs—and even heavy local Vision-Language Models—meant building a deterministic C++17/Rust layout parser (ocr_engine) from scratch so the local LLM simply reasons over structured text without ever touching raw pixels: Tabular Balance Sheets (Production): Accelerated with explicit SIMD dispatch (SSE4.2, AVX2, AVX-512F), parsing images in 30–200 ms (depending on input resolution/density) within a 10–30 MB memory footprint. Chart OCR (Dev Branch): Currently running purely scalar at ≤600 ms max within a ~40 MB footprint—with significant headroom still left once SIMD vectorization lands there too. Pairing that with a zero-allocation C23/Rust voice pipeline for Hindi/English prompting keeps the entire loop fast, private, and completely offline by default. Glad the opt-in sync philosophy resonated—it’s how personal finance software should have been built from day one.