Reading handwriting off a wall of sticky notes
Plenty of people plan things on a wall. A wedding, a renovation, a term's coursework, a sprint. Sticky notes on a board, moved around by hand, because it is genuinely a good way to think.
The problem is that the wall does not leave the room. You cannot search it, you cannot share it with someone across town, and if it falls down or the room gets cleared, the plan is gone.
WallMind photographs the wall and turns each note into text you can search, share and back up. The physical wall stays exactly as it is — that is the point. We are not replacing it.
Two problems, not one
"Read the wall" is really two jobs, and they fail differently.
Detection — finding the notes. Where does each square of colour begin and end, at an angle, in whatever light the room has, possibly overlapping its neighbour.
Recognition — reading what is written on each one. This is the hard half. Printed text recognition is largely a solved problem; handwriting is not. Real notes are written fast, in marker, by people with no interest in legibility. A word your own brain reads instantly from context can be genuinely ambiguous in isolation.
Where the reading happens
This was the decision worth thinking about, and it is a privacy question disguised as an accuracy question.
The accurate answer is to send every photo to a large model on a server and let it do the reading. Better results, no question.
By default, WallMind does the reading on your device, using Google ML Kit. Everything happens locally. It works with no network connection. No image of your wall is transmitted anywhere.
That is a real accuracy cost, and we took it, because the default behaviour of an app is the behaviour most people will ever experience. A default that quietly uploads photographs of whatever is written on your wall — which might be your finances, your medication schedule, or your family's plans — is not a default we wanted to ship.
The escape hatch, and whose key it uses
Some walls are genuinely hard to read. So there is an opt-in: you can add your own Google AI Studio (Gemini) API key in Settings, and scanned images are then sent to Gemini for better recognition.
The detail that matters is whose key. It is yours, not ours.
- The request runs under your account, on your quota, subject to your agreement with Google — not a pipeline we operate on your behalf
- It only happens while a key is set. Remove the key and the app returns to reading everything locally
- Nothing is silently upgraded. Better accuracy is a thing you switch on deliberately
We could have put our own key in the app and made it the default. It would have looked better in a demo. It would also mean every scan flowed through us, which is precisely what we were avoiding.
Making imperfect recognition acceptable
If recognition is going to be imperfect — and on handwriting it always is — the answer is not to chase perfection. It is to make correction cheap.
So the scan produces a review step: the detected notes and the text read from them, before anything is kept. Fix what is wrong, then keep it. A misread word costs a tap.
This changes what "good enough" means. On-device recognition that gets most of it right, with a fast correction pass, beats a perfect cloud pipeline that costs you your privacy — for this particular job, where a human is standing right there anyway.
What happens to the photograph afterwards
The captured image has served its purpose once the notes are extracted. Keeping it forever is storage you did not ask to spend.
So wall photos are deleted from the device automatically after 30 days. The searchable text — the part with the value — stays.
If you want the images kept, there is an optional Drive backup: full-size photos are copied into a WallMind folder in your own Google Drive, filed by month, so they outlive the cleanup on your phone. As with the Gemini key, it is off until you turn it on, and it goes to your Drive rather than to us.
There is one deliberate ordering detail: a photo is only deleted from the device once it has safely reached your Drive. Cleanup that runs ahead of backup is just data loss with extra steps.
The through-line
Every one of these is the same decision made three times: do the private thing by default, and make the more capable thing an explicit choice the user makes with their own account.
It costs some accuracy in the default path and it costs a settings screen. What it buys is that the honest description of the app — everything happens on your phone unless you decide otherwise — is a statement about how it works, not a promise about how we behave.