Veil

Veil redacts video from a plain-English instruction, then checks its own work. Anyone who films in public ends up with footage full of people who never agreed to be in it. Journalists, creators, schools, event organisers and agencies answering bodycam or CCTV records requests all face the same choice: blur everything automatically, or mask by hand, frame by frame. Neither handles what people actually need ("everyone except her", "the plate on that car", "also the phone"), and no tool tells you when it missed someone. With Veil you type the instruction: "Blur every face except the woman in the pink coat." Llama 3.1 8B turns it into a validated plan of tool calls (segment, select, subtract, union, effect). SAM 3.1 tracks every face, person and coat, and per-frame mask logic applies the exception. Then SAM 3.1 re-segments every frame of Veil's own output: anything still visible is re-prompted at the leak, tracked and patched, and whatever can't be fixed is flagged with a timestamp for a person to review. Follow-up edits ("Also hide the bank logo") reuse earlier tracks and segment only the new concept. Every run ends with a Redaction Report. The original, the plan, the redacted video, the masks and the audit log are chained with SHA-256 hashes, so changing a single byte fails verification. On a 13-second Manchester street clip, the audit re-checked all 397 output frames, patched 19 misses from the first pass and flagged 12 frames for review. The whole pipeline runs on a free Colab T4. The live demo replays the recorded runs instantly, no GPU needed: https://atharva283--veil-demo-web.us-east.modal.run