CJ-01 · 3U
MANTRANSLATED
AI tool that reads the text off manga, manhwa and manhua pages and turns a whole chapter into a bilingual .txt file.
- Built by
- Chananon Jankaew
- Stack
- Python · FastAPI · Gemini · Cloud Run · Cloud Storage · Firebase
The problem
Reading a chapter in Chinese, Korean or Japanese means pulling the text off every page before you can translate it, and a chapter can run to a hundred images or more. The tool had to take a whole chapter at once and keep the speech bubbles in reading order.
Result
- Caching the shared prompt cut Gemini input cost by about 40%.
- Takes up to 200 images or 50 MB in a single ZIP.
- The output pairs each original line with its translation, ready to read side by side.
How it's built
- 01One multimodal Gemini call per page reads the speech bubbles in order and translates them in the same request, instead of a separate OCR step followed by a translation step.
- 02The browser unzips and resizes the chapter, then uploads it straight to Google Cloud Storage through a resumable session, so large chapters never hit the API's request-size limit.
- 03Pages run eight at a time with exponential-backoff retries on 429 and 5xx errors, yet the output always follows page order; progress streams back to the browser as NDJSON.
- 04On Cloud Run each container takes one job and the service scales out to ten; Firebase Auth and Firestore handle accounts and per-image quotas.
