Picture guard

People paste screenshots, payslips, letters and ID photos into AI chatbots every day. Privacy tools that check typed text cannot see the words inside a picture. Picture guard closes that gap. It finds names, addresses, account numbers, ID numbers and faces in a picture, covers each one with a black box, and sends only that cleaned copy to the AI. If it is not sure, it blacks out the whole picture. How it works: Tesseract reads every word and its position. Muse Glimmer, Meta's open-weights model, running on our own GPU, decides which words are private. Pattern rules with check digits catch account, card, IBAN and National Insurance numbers. SAM 3.1 finds faces. Picture guard then reads its own cleaned copy again; if anything private is still readable, the whole picture goes black. Only then does Muse Spark, through the Meta Model API, answer the user's question — from the cleaned copy only. The page proves it: the fingerprint of the bytes sent to Meta matches the cleaned copy, never the original. Results on 90 sealed test pictures it had never seen (run once, on a frozen build): 0 private values left readable by three separate checks; 26% of text pictures blacked out on purpose by the safety rule; 78 of 79 faces covered at 90% or more in the pictures it kept (one hard-face picture was blacked out whole). Limits are stated in LIMITS.md: made-up test pictures and AI-made faces only. Built by navee4 (solo builder) for the hackathon, on top of a test-picture generator and model experiments made on 20–21 October (listed in BUILT-BEFORE-HACKATHON.md; the pipeline itself was built after 26 October). AI coding tools (Claude) helped write the code. The demo video's voice-over is an AI voice (ElevenLabs, elevenlabs.io) reading a script by navee4. Open source, Apache-2.0.