arinais-stuborn ♡

Turn lecture slide PDFs into a page you can import into Notion, with the same wording and the same diagrams. Everything runs in the browser: no AI, no server, and the PDF never leaves the visitor's device.

What it does

1. Upload a PDF (drag & drop or file picker).
2. The converter reads every slide with pdf.js:
   • text with fonts, sizes, colours and positions
   • vector shapes (boxes, ovals, arrows) and embedded images
3. Template detection: anything that repeats in the page margins on most pages (logo, author footer, page numbers) is ignored.
4. Per slide:
   • text outside graphics → real Markdown: ## heading, bullets, sub-bullets, numbered lists, → lines and tables. Bold, red/coloured and underlined words become bold.
   • graphics (and the text inside them) → one cropped PNG, with logo and footer painted out.
   • vertical side notes ("Quelle: …") → a text line under the slide.
   • empty slides → *(Folie N – leer)*.
5. You get back <name>_Notion.zip containing <name>.md and an images/ folder. In Notion, go to Import → Markdown & CSV and pick the zip.

Notes → cute PDF (second tab, #notes)

1. Scan or pick photos of handwritten notes (the camera button opens the phone camera), or just type.
2. The handwriting and maths are read by GLM-OCR (0.9B, MIT). Formulas come back as LaTeX ($…$ / $$…$$). Each photo is scaled to 1000 px wide and read in strips of ≤ 0.4 MP, cut at the emptiest rows (blank strips are skipped), which keeps memory small and each step short.
   • Computers: on the device. GPU via Transformers.js 4 + WebGPU (onnx-community/GLM-OCR-ONNX, 4-bit). The model (~800 MB) comes from Hugging Face's CDN the first time and is cached by the browser; it's freed after each scan and reloads from the cache. A download that gets no bytes for 45 s counts as failed; the next scan resumes it. (It can't live on Vercel: Hobby uploads are capped at 100 MB per file and functions have no GPU.)
   • Phones: the reading server. A phone tab can't hold the model (iOS Safari kills the tab), so phones (and computers without WebGPU, or whose on-device reader fails) send the strips to SERVER in notes.js: the same model (ggml-org/GLM-OCR-GGUF Q8_0) behind llama.cpp's OpenAI-compatible API on Modal, see modal-reader/app.py. 8 CPU cores, a few seconds a strip; it scales to zero when idle and wakes in ~15 s (the page waits up to 5 min). At most one container runs. Nothing is stored. Runs on Modal's monthly free credits; the account has no payment method, so it can't bill. Redeploy with python -m modal deploy modal-reader/app.py.
   • Last resort: Tesseract.js 5 (English + German, text only) when the server doesn't answer. A note in the editor says so.
3. She fixes anything it misread in a small markdown editor: # title, ## heading, - bullet, 1. number, [ ] to-do, > sticky note, **highlight**, $x^2$, $$ formula $$ (a line that is a single $…$ formula counts as a block).
4. She picks a paper (strawberry milk, lavender dream, peach blossom) and a handwriting font (Caveat, Gaegu, Patrick Hand). A live preview shows the A4 pages, with washi tape, doodles, a motivational sticker and a footer on each one. Formulas are typeset by MathJax 3 (loaded only when there is maths); its SVG paths are drawn straight onto the canvas in the ink colour, block formulas on a soft card.
5. Downloads:
   • <title> ♡.pdf, made with jsPDF
   • <title>_Notion.zip: the editable text (formulas as $…$ / $$ blocks) plus the cute pages as images/page-NN.jpg, imported the same way as the slides zip

Project layout

public/              ← the website (static, no build step)
  index.html
  css/style.css
  js/app.js          ← UI: upload, progress, results, download
  js/converter.js    ← the deterministic PDF → Notion converter
  js/notes.js        ← notes tab: on-device handwriting + maths reading, cute pages, PDF + Notion zip
  vendor/            ← pdf.js 4.10 + JSZip 3.10 (self-hosted, no CDN needed)
                       (the notes tab loads tesseract.js 5.1.1, jsPDF 2.5.2, Transformers.js 4.3.1 and
                        MathJax 3.2.2 from jsDelivr, and the GLM-OCR model from Hugging Face)
tests/
  harness.html       ← runs the converter on tests/…/sample.pdf
  run_harness.py     ← headless-Chrome test: python tests/run_harness.py slides.pdf out/
modal-reader/app.py  ← the reading server for phones (Modal: llama.cpp + GLM-OCR)
vercel.json          ← static hosting config (serves public/)

Run locally

python3 -m http.server 8080 -d public
# open http://localhost:8080

(Any static file server works. Opening index.html straight from disk does not work, because ES modules and web workers need http.)

Deploy to Vercel

Option A: GitHub (recommended)

1. Push this folder to a GitHub repo.
2. On vercel.com, click Add New → Project, then import the repo.
3. Leave everything on default. vercel.json already sets Framework: Other and Output directory: public, with no build command.
4. Click Deploy.

Option B: Vercel CLI

npm i -g vercel
vercel          # preview deployment
vercel --prod   # production

Licences

pdf.js: Apache-2.0 (see public/vendor/PDFJS-LICENSE). JSZip: MIT. Tesseract.js: Apache-2.0. jsPDF: MIT. Transformers.js: Apache-2.0. GLM-OCR: MIT. MathJax: Apache-2.0. Fonts: SIL Open Font License (Google Fonts).