Blog
-
Letting Users Fix What OCR Missed — and What Their Corrections Taught Us
A detector that misses text produces no score, no count, and no trace — so we let users point at it. Four attempts at the interface, including the one that shipped completely broken on mobile for a CSS reason worth knowing, and the two pipeline changes the first seventeen corrections produced.
-
How OCR Actually Works: Detection, Recognition, and Where It Fails
OCR is two models, not one, and they fail in opposite ways — one goes silent, the other invents. Why average confidence is the wrong metric, why the detection canvas size mattered more than every threshold we tuned, and what 17 user corrections revealed about the misses no metric catches.
-
Vertical Japanese Is the Hardest Case for OCR — Here's Why
Vertical writing breaks five assumptions built into standard OCR pipelines: line shape, reading order, crop rotation, detection granularity, and rendering. Including the exception in our own code that was documented, tuned, and had never once fired on the text it was written for.
-
Translating Text Inside an Image Without Wrecking the Layout
Reading and translating are the easy parts. Putting the translation back means erasing the original without destroying its background, recovering the ink colour, refitting text that changed length, and picking a font that can draw every character — including the ones the translator left behind.
-
Removing Text From an Image: What Inpainting Gets Right and Wrong
Generative inpainting is the right tool for text over a photograph and the wrong one for a red band inside a white panel — where it obediently erases the band. How to tell the cases apart, and how to erase glyphs without touching the label they sit on.
-
Pulling the Same Field Out of a Stack of Documents
Invoices from one supplier share a layout, so the total is in the same place every time. Using position as the signal, why regions are stored relative to the page, where the approach breaks, and what batching ten documents demands that running one does not.
-
Compositing a Product Into a Photo, Entirely in the Browser
Cutout, tone match, contact shadow — the three corrections that separate a composite from a sticker. Why running it client-side changes the design more than the output, why placement stays manual, and what still needs a studio.