Developer Turns reMarkable Paper Pro into AI Diary
Maxime Rivest's open-source riddle project turns the reMarkable Paper Pro into a Tom Riddle-style AI diary that reads handwritten prompts and writes replies back on the E-Ink screen. The GitHub repo says the mod commits the page as a PNG, sends it to a vision-capable LLM backend, and streams the response sentence by sentence; Android Authority reports replies can begin appearing in roughly one second. For practitioners, the useful pattern is not the novelty theme but the UI pipeline: handwriting capture, image-to-LLM inference, and stroke-by-stroke rendering that hides latency on a low-refresh display. The project is also a cautionary prototype, because it requires developer mode, SSH access, root-level device changes, and an OpenAI-compatible API or local image-capable server.
Handwriting-first AI interfaces usually fail on latency and interaction feel, so the useful signal here is the complete pipeline: capture ink, send a page image to a vision-capable model, and render the answer as animated strokes quickly enough that the device still feels like paper rather than a chat app. That makes this a prototype worth studying for teams building embedded, tactile, or low-refresh AI experiences, even though it is not a consumer-ready product.
What happened
Maxime Rivest published riddle, an open-source project for the reMarkable Paper Pro that turns the E-Ink tablet into a Tom Riddle-style AI diary. The GitHub repository says the app captures pen strokes, waits for the page to settle, commits the handwritten page as a PNG, sends it to a resident LLM process, and streams the reply back sentence by sentence. Android Authority reports the response can begin appearing in roughly one second and describes the reply as animated handwriting on the E-Ink display.
Technical context
The implementation matters because it is not just a prompt wrapper. The repository describes Rust pen-input handling, handwriting synthesis, a display backend, and a vision LLM backend that can use an OpenAI-compatible API, OpenRouter, Groq, or a compatible local server that accepts image input. In practice, the pattern is handwriting capture plus image-to-text reasoning plus incremental rendering. The animated writing hides some generation delay while preserving the paper metaphor.
For practitioners
The project is a good reference for product teams exploring non-chat AI interfaces, especially on devices where keyboards and conventional chat windows are poor fits. It also exposes the engineering cost: the repo says installation needs developer mode, SSH, AppLoad or related tooling, and in takeover mode root-level control of the display stack. Any production version would need a much cleaner permission, recovery, and privacy model before routing handwritten notes to cloud APIs.
What to watch
Watch whether the project becomes a reusable pattern for notebooks, field-service devices, accessibility tools, or education demos. The strongest follow-up evidence would be a safer install path, clearer privacy controls for handwritten pages, and examples that use local vision models rather than only cloud-compatible endpoints.
Key Points
- 1The demo proves a handwriting-first LLM interface can feel responsive when image capture, streaming generation, and stroke rendering are coordinated.
- 2GitHub's implementation shows device-level work matters: developer mode, root access, AppLoad, and display control are required.
- 3The architecture shifts compute to OpenAI-compatible or local vision backends, creating privacy and reliability questions for copied designs.
Scoring Rationale
This is a useful proof-of-concept for alternative human-AI interfaces and embedded LLM UX, especially around handwriting capture and streaming display. Its impact remains moderate because it is a niche open-source device mod rather than a broadly deployed platform or model advance.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
