AI proofreading that stays on your device.
On-device AI flags errors in your PDFs — without uploading your documents anywhere. The AI suggests. You decide.
How it works
Apple’s Foundation Models, on your iPad.
The text from your PDF is processed locally — on the hardware in your hands — and the results appear as suggestions you can review, accept, or dismiss.
No internet connection required. No data sent to servers. The AI scans for spelling errors, grammar issues, and inconsistencies. When it finds something, it flags the location. You decide: stamp it or dismiss it.
AI proofreading screenshot
Your clients’ documents aren’t training data.
This isn’t a marketing claim. It’s architecture. On-device means on-device.
Nothing uploaded
The text never leaves your iPad. Not temporarily. Not “encrypted.” Not at all.
Nothing retained
There’s no server to retain anything on.
Nothing shared
Your documents can’t train models because they never reach one that isn’t on your device.
What the AI does well
Spelling errors and typos, including context-sensitive ones
Grammar: subject-verb agreement, tense, commas
Formatting inconsistencies and spacing anomalies
Names with different spellings, inconsistent numbers
Best as a safety net for the errors that slip past tired eyes.
Honest limitations
Won’t catch every error. No AI will.
Doesn’t understand house style guides beyond common conventions.
May flag intentional stylistic choices as errors.
Limited on highly technical or domain-specific terminology.
That’s why you review every suggestion. Your expertise is the final word.
Optional
Cloud AI models, when you choose.
For documents where you need deeper analysis, optional cloud AI models provide more sophisticated analysis than on-device models alone. Never automatic — you choose to enable it per document, you choose which model, and you see clearly which documents use cloud processing.
“AI proofreading” sounds like a gimmick. We understand.
Here’s what it is, plainly: spell-check and grammar-check that’s significantly smarter than word processors, running locally on hardware that’s finally powerful enough to do it well. Pattern recognition applied to text — good enough to be useful as a first-pass filter.
If you’ve ever finished proofreading a 300-page manuscript and then found a typo on page 4 that you swear wasn’t there before — that’s the problem this solves.
Use it or don’t. Stampede works perfectly well without it.