scrolloom™ — AI Writing Diagnostics

Don’t just ask “AI or human?” Ask what the evidence says.

Drop a text-based PDF. The app reads every page locally, runs document-wide writing diagnostics, and can optionally run a real on-device AI-text classifier over the entire extracted text. Nothing in the PDF is sent to a server.

↥

Drop a PDF here

or choose one from this device.
Text-based PDFs work best; image-only scans need OCR and are reported as unreadable.

Analysis settings

Deep local model scanUses TMR RoBERTa AI-text detector locally in your browser. First run downloads an ~126 MB quantized model. The browser caches it when Cache Storage is available.

Explainable writing diagnosticsRhythm, repetition, paragraph regularity, formulaic discourse, lexical diversity, structural uniformity and PDF provenance signals.

Document language The classifier is primarily trained for English. For non-English documents the app will still run the full explainable scan, but will not pretend the English model is authoritative.

Important: AI-text detection is probabilistic evidence, not proof of authorship. This tool is intentionally designed to show uncertainty, coverage and disagreements instead of forcing every document into “AI” or “human”.

Preparing…

—

AI-signal score

Analysis complete

Model evidence

Local classifier results across all analyzed chunks.

Why the writing was flagged

Independent style signals. These explain patterns; they do not by themselves prove AI authorship.

Page heatmap

Every page with extractable text is included. Tap a page for its evidence.

Document & provenance

Extraction coverage, metadata and invisible-character checks.

The deep model used here is the open TMR RoBERTa detector trained on the RAID dataset and is primarily calibrated for English. Model scores can be wrong, particularly for short, unusual, heavily edited, translated, technical or out-of-distribution text. The app therefore reports evidence, coverage and confidence rather than claiming authorship verification.