I built an OCR service for Japanese Kindle books so I could feed them to AI without hallucinations
I read a lot of Japanese non-fiction on Kindle. When I started using AI (ChatGPT, Claude) to discuss what I'd read, I hit a wall: the AI would confidently make up quotes, mix up arguments, and hallucinate details. The only fix is giving it the actual text — but Japanese Kindle books have no usable export. Highlights are limited and copy-paste is locked down.
So I built BookHalo. You upload screenshots of your Kindle pages, it OCRs them with high accuracy for Japanese text, and gives you clean text you can paste straight into any AI chat. The difference between "AI summarize this book" with and without the actual source text is night and day.
Still early — running it as a side project alongside a day job. The OCR accuracy for Japanese vertical text was the hardest part to get right.
Curious if anyone else has run into this problem with non-English books, or if there are workarounds I'm missing.
Feeding AI the actual source text versus asking it to recall from training is basically using two different tools — the quality gap is larger than most people expect.
I ran into the same extraction pattern building a YouTube caption bot. The platform makes automated access intentionally difficult, the workarounds break regularly, but once clean text is out the AI response quality jumps immediately. The extraction problem ends up being more work than the AI part.
Japanese vertical text OCR on top of Kindle's copy-paste restriction is a hard stack to crack. The screenshot path makes sense when there's nothing else to work with.
Yeah, the "workarounds break regularly" part is painfully real. Totally agree there. And on the Japanese vertical text OCR — it was genuinely the hardest part, but we got it to a point where the accuracy is honestly terrifying now. Turns out once you crack that, the rest of the pipeline is almost boring by comparison.