When the free tier died, my OCR moved into the browser

Klartext reads printed documents — photo in, plain text out, with the preprocessing pipeline made visible. I built it the respectable way first: FastAPI, Tesseract and OpenCV in a Docker container, tested end to end locally, ready for a free Hugging Face Space.

Then I opened the Space creation page and found a new banner: Docker Spaces now require a paid plan. My deployment target had vanished between design and deploy.

The rewrite that was an upgrade

The forced question — where else can inference run for free? — had an answer hiding in plain sight: on the user's machine. Tesseract compiles to WebAssembly, so the recognition engine can run inside the browser tab itself. I rewrote the pipeline client-side: preprocessing on a <canvas> (grayscale, then Otsu binarization implemented from scratch — it's thirty lines and worth understanding), then tesseract.js for recognition with per-word confidences.

What started as a workaround turned out to be the better architecture:

• Privacy stopped being a promise and became a property. Your document isn't "handled securely" — it never leaves your device, because there is no server to send it to.
• No cold starts, no hosting bill, no ops. A static page does inference.
• The tested Docker container didn't die — it lives in the repo as a documented self-host option, with the deeper OpenCV pipeline (denoising, automatic deskew) that the browser version doesn't need for its job.

What I actually learned

The OCR lesson: preprocessing dominates. The same engine that shrugs at a raw phone photo reads near-perfectly after binarization — which is why Klartext shows you every stage instead of hiding them behind a spinner.

The engineering lesson is the one I'll retell in interviews: dependencies on someone else's pricing page are architecture decisions too. When the ground moves, the ability to re-plan — and to recognize when the workaround is actually the better design — matters more than the original plan ever did.