Zero-Retention Context Engine
On-device macOS engine that understands the screen while provably retaining no pixels — MobileCLIP embeddings on the Apple Neural Engine, privacy adversarially verified.
Overview
A context engine that watches the screen via ScreenCaptureKit, reads it with on-device Vision OCR, embeds frames with MobileCLIP via Core ML, and keeps only L2-normalized vectors — the meaning without the pixels. The privacy claim is adversarially verified, not just asserted.
Problem
Screen-context assistants are a privacy nightmare: they either stream your screen to the cloud or silently retain screenshots. Can a system remember what you saw without keeping anything a human could look at?
Approach
Frames are embedded on the Apple Neural Engine and discarded in memory — measured at 3.0ms median per frame on M1 (329 img/s) and 2.0ms on M5, ~9.6–11.8× faster than CPU, with export correctness verified against the PyTorch reference (0.9965 cosine). Built entirely on Apple's on-device stack: ScreenCaptureKit, Vision, MobileCLIP, Core ML.
Impact
Rigorous on-device ML engineering: hardware-aware benchmarking, verified model export, and a falsifiable privacy claim — systems thinking applied to the most sensitive data there is.