How CloakMind is built

What it is

CloakMind Tape is a small, single-purpose device for working through printed exercises: past papers, problem sets, textbook questions. You point it at one exercise on the page and press the screen. It shows you the text it read, you check that against the paper, and then it walks you through the solution on its own display.

It is not a phone app and not a browser extension. It has its own camera, its own screen and its own battery, and it does nothing else — no notifications, no feed, nothing to switch to while you wait.

The device is prelaunch. No unit has shipped, the enclosure and battery are not final, and the waitlist is the only way in. The planned founder price is 99,90 € against a planned retail price of 119,90 €, with the AI billed separately from 0 € if you bring your own key. Those are current plans, not a binding offer.

What it does not do

The model does not run on the device. The device captures a photo, uploads it over Wi-Fi and renders the result; the transcription and the reasoning happen on servers. So a connection is required, there is no offline mode, and a full pass takes about 30 seconds.

It also cannot tell when it has misread your page. That is why the transcription is shown before the worked answer rather than after: a misread exponent produces a flawless solution to a different exercise, and nothing inside that solution will look wrong. Checking the transcription against the paper is part of using it.

The hardware

The prototype is built on a Waveshare ESP32-S3 touch-LCD board: a 2 inch capacitive touchscreen, 16 MB of flash and octal PSRAM, with an OV5640 5 MP autofocus camera module on the board's camera header.

The ESP32-S3 was the right part for a reason that is not performance: it has a native USB peripheral and enough PSRAM to hold a JPEG frame buffer and an LVGL draw buffer at the same time. A cheaper part would have forced a choice between the camera resolution and the interface.

Firmware is Arduino-framework on espressif32 7.1.0, built with PlatformIO, with a dual-OTA partition layout on the 16 MB flash so an update never overwrites the running image.

The firmware

The interface is LVGL. Every screen is a mode in one state machine, and anything that blocks — a capture, an upload, a settings write — is queued as a pending action rather than run inside a touch callback, so the UI never stalls on the network.

The whole firmware is a single translation unit. That is a deliberate constraint rather than an accident: one file that has to stay readable is harder to let sprawl than twelve that each look small.

Fonts and icons are not linked in as a library. They are generated from source SVGs and TTFs into C arrays by a small tool in the repo, and the generated headers are committed, so a build never depends on a font being installed on the machine doing the building.

Updates and recovery

Over-the-air updates are signed. The release tool builds a manifest describing the image and signs it with an RSA-3072 key held outside the repository; the matching public key is compiled into the firmware, and the device verifies the manifest signature before it writes a single byte to the inactive OTA slot. An unsigned or mismatched image is refused on the device, not on the server.

If a device will not start at all, there is a USB recovery page on this site that flashes it over Web Serial straight from Chrome or Edge — no toolchain, no drivers, no command line. Recovery rewrites the partition table, the OTA selection data and the signed application, and deliberately leaves the bootloader and the saved device settings alone.

That path exists because the failure it covers is the one that matters: a bad update on hardware that is already in somebody's hands, far away from anyone who can open a terminal.

What is still unfinished

The enclosure. The current one is printed and it is a prototype in the honest sense of the word, not the marketing one.

Battery life, which has not been characterised across a real day of use, and thermals under sustained capture, which are still being measured.

Handwriting. Printed exercises work considerably better than handwritten ones, and closing that gap is a model and capture problem that is not solved yet.

And the economics of the AI plans. The prices on this site are the current plan and they have not been tested against real usage at any volume.

Getting in touch

The firmware and the tooling around it will be published under a permissive licence; this page will link to the repository once it is public, rather than link to something that does not exist yet.

Questions about the hardware, the signing scheme or anything else on this page go to [email protected]. Corrections are welcome and get made.