beyond

MirrorCheck — AI-Assisted Impersonation Review MirrorCheck turns a suspected copied social profile into evidence a human can review. Creators and ordinary users can report an apparent account copy from an Instagram-style web interface, attach original and suspected-profile screenshots, and follow the decision through to a visible profile warning. PaddleOCR extracts editable account fields; imagededup finds reused photo regions. A local OCR + Microsoft Presidio pipeline masks detected email and phone regions before any screenshot reaches Meta. Failed redaction blocks external analysis. Meta Muse Spark 1.3 returns structured visual similarities, differences, limitations and a constrained recommendation. Separate local models show profile-pattern and pair text-reuse signals. A standard avatar is excluded from the profile model's features. AI cannot add a warning. A moderator must inspect the processed evidence and record an explained decision. Confirmed reports display a warning; dismissed reports do not. Version 0.13 adds audit history, report-scoped appeals, SQLite persistence and isolated visitor sandboxes. Built with React, TypeScript, Vite and FastAPI, with authenticated local workers and server-side Meta credentials. The public demo runs the real API and local ML in a US West Codespace. Each visitor gets a fictional sandbox; only the web port is public. A built-in example pair supports the full review without uploading private files. Temporary live demo (Codespace must be running): https://congenial-garbanzo-4j5v4xwwx9r3jw9v-8090.app.github.dev/demo-guide.html Evaluation separates copy detection from spam patterns. Controlled external-text copy recall improved from 62/100 to 92/100; a difficult Copydays photo-region test improved from 1/50 to 18/50. These are component evaluations, not real Instagram impersonation accuracy. Pair-model training is synthetic; no screenshot or score establishes ownership. This is an unofficial feature prototype, not an integration with the real Instagram service. The recorded three-minute English demo uses fictional accounts, edited authentic browser captures and synthesized narration. Codespaces is temporary attended hosting; the independently hosted video remains available when the live app sleeps. Source, model artifacts, evaluation methods and third-party notices are public. Future work includes consented account-level validation and verified ownership evidence. Evaluation sources and limitations are documented in the repository. No raw external profiles are redistributed.